2d:null 2e:["$","title",null,{"children":"The Role of Contrastive Learning in Visualization | Pezanne Khambatta | Observable"}] 33:T645,md`## Research Paper Details The paper I'll be discussing is detailed below: ### Title Supporting Analysis of Dimensionality Reduction Results with Contrastive Learning ### Links Publication Link: [https://arxiv.org/pdf/1905.03911.pdf](https://arxiv.org/pdf/1905.03911.pdf) Published In: [https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=2945](https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=2945) Official IEEE link: [https://ieeexplore.ieee.org/document/8805461](https://ieeexplore.ieee.org/document/8805461) ### Authors [Takanori Fujiwara](https://ieeexplore.ieee.org/author/37086133964), [Oh-Hyun Kwon](https://ieeexplore.ieee.org/author/37085484504) and [Kwan-Liu Ma](https://ieeexplore.ieee.org/author/37085432740) ### Publish Date 15th Oct, 2019 ### Description When analyzing a dataset with several attributes, we generally make a choice of selecting a handful of these attributes and ignoring the others based on trial-and-error prototyping and some level of instinct. However, the publication would suggest that this is a naive approach of tackling high-dimensional data. One solution is to identify *clusters* of data points and determine which features of the dataset are driving the variance or *contrast* in these clusters. One way of doing so is identified to be **Contrastive Clusters in Principal Component Analysis** (ccPCA) that relies on the priciple of contrastive learning and aims to discover patterns that are specific to, or enriched in, one dataset relative to another. Now, we have made a much more educated guess at which features to focus on. `34:T436,md`## How Research was Conducted

${Object.assign(await FileAttachment("workflow.png").image(), {style: 'width: 100%; max-width: 550px;' })}

Fig. 1: The Analysis Workflow

The workflow for analyzing high dimensional data using DR is broken down as follows: **a.** Applying a DR method on high-dimensional data **b.** Identify clusters in the DR result by applying a clustering method. Afterward, the task is to understand the clusters’ characteristics. **c.** Finding features (or dimensions) which have a high contribution to contrasting each cluster with the others. **For this step, we utilize cPCA.** **d.** Reviewing the detailed differences of values of the highly contributed features between each corresponding cluster and the other data points. Steps *c.* and *d.* help us obtain an understanding of which and how features contribute to the uniqueness of each cluster. Now, the user can update the DR result or clusters by selecting a subset of the data points based on their interest, as illustrated above. `35:Tb11,md`## Visualizations Demonstrated The publication is rich with visualizations as shown below. ### 1. Collection - Screenshots of Researchers' Prototype System UI

${Object.assign(await FileAttachment("workflowEg.png").image(), {style: 'width: 100%; max-width: 600px;' })}

Fig. 2: The dimensionality reduction (DR) view (a) visualizes a result after DR and clustering. The feature contributions view (b) shows the measures of each feature’s contribution to contrasting each cluster with the others. The feature values of the selected cells in (b) are visualized as histograms, as shown in (c). In (d), the user can change the settings for the analysis methods and visualizations.

${Object.assign(await FileAttachment("workflowEg2.png").image(), {style: 'width: 100%; max-width: 600px;' })}

Fig. 3: An analysis result of female players from the Tennis Major Tournament Match Statistics dataset.

${Object.assign(await FileAttachment("workflowEg3.png").image(), {style: 'width: 100%; max-width: 600px;' })}

Fig. 4: A result of the Nutrient dataset. (a) shows the result after applying t-SNE and DBSCAN. A point’s color and size show the clustering label and the value of ‘calories’, respectively. (b) shows the FCs of each cluster. (c) shows the histograms of the selected cells in (b), as indicated with the colored numbers in both (b) and (c).

${Object.assign(await FileAttachment("workflowEg4.png").image(), {style: 'width: 100%; max-width: 600px;' })}

Fig. 5: A screenshot of researchers' prototype system. The dimensionality reduction (DR) view (a) visualizes a result after DR and clustering. The feature contributions view (b) shows the measures of each feature’s contribution to contrasting each cluster with the others. The feature values of the selected cells in (b) are visualized as histograms, as shown in (c). In (d), the user can change the settings for the analysis methods and visualizations.

#### Visualizations The visualizations (UI, rather) are a collection of small visualizations, all illustrating the clustering of different datapoints. These small visualizations include a scatterplot, bar charts and a heatmap. #### Encoding All of the above visualizations are encoded in their traditional sense. Color is used to differenciate different clusters. Color is also used in the heatmaps to represent user-defined value used to perform the classification. #### Data-Ink Ratio The data-ink ratio of the UI is very high. Several visualizations exist and each of them have a lot of data. #### Data Density Much like data-ink ratio, this is high. #### Lie Factor The visualizations don't have much of a lie factor. `36:T661,md` ### 2. Classifying Wine from Chemical Properties

Fig. 6: Parallel coordinates showing all features in the Wine Recognition dataset. The corresponding polylines for the wines are highlighted with (a) black, (b) green, (c) orange, and (d) brown clusters. It is difficult to discern the essential features from this visualization.

#### Visualization The visualization is a small-multiples parallel coordinates chart with 4 separate charts stacked vertically. Each chart is identical except for the fact that the respective identified cluster is highlighted in its own color. Color encodes cluster. #### Encoding the y-axis of each small chart depicts normalized value from 0.0 to 0.1, and the x-axes each depict discrete values of 13 different features of wine. #### Data-Ink Ratio The data-ink ratio is very high because a lot of data points have been shown in a small space. To eliminate redundancy of lines, the small-multiples could be replaced by a single chart, thus making it easier to tell the difference between each classification. #### Data Density Much like data-ink ratio, this is high. #### Lie Factor The only lie factor is that the color v/s non-color over-emphasizes the classification and overpowers the visual aspect of detecting a cluster just positionally. From the chart, it is hard to visually be convinced that the clusters have been identified correctly - which is a good thing because that is what the research paper wants to illustrate. ` 37:T783,md` ### 3. Classifying Digits form Handwriting

Fig. 7: Comparison of features’ relative contributions of MNIST digits. Researchers compare LDA, PCA, and ccPCA. All of these methods can calculate the features’ relative contributions to the first component by respectively referring to either LDA’s loadings, PC loadings, or cPC loadings de- scribed in Sect. 4.3. We scale each loading in the range from -1 to 1 by dividing the maximum absolute value of the loadings. We visualize the scaled loading for each pixel with a blue-to-red colormap.

#### Visualization The visualization is a matrix of heatmaps. #### Encoding The x-axis of the matrix depicts the clustering method used and the y-axix depicts the digit being identified. The row and columns in each cell of the matrix the cartesian position of the features picked up by the handwritten text. The color-scale is a diverging gradient ranging from -1 to 1, representing the feature's contribution (normalized). #### Data-Ink Ratio The data-ink ratio is moderate to high. It is selectively high for the PCA and cPCA columns but very low for LDA. This is understandable for the sake of illustrating the point that LDA is not so good at classification. #### Data-Density Much like data-ink ratio, the data density is moderate to high. #### Lie Factor it would be tempting to say that using a diverging gradient color scale usually overemphasizes variance, but in the case of a classification visualization, it has been used appropriately. The lie factor is low. From the visualization, it is visible that ccPCA is the most effective method to calculate each feature's relative contribution, judging by the fact that the digits look more pronounced for the same. PCA is a close second and LDA is left far behind. `38:T694,md`## Insight & Opinion ### Relevance cPCA can be used to find features contrasting the target cluster. It is a linear method of clustering as opposed to many of the nonlinear DR methods used for visualization that generate irreversible low-dimensional projection of the original data structure. These methods do not have a parametric mapping between the original and projected dimensions, thus making it difficult to provide information about how these DR methods affect cluster forming. cPCA provides flexibility for analyzing results from any type of DR methods. This is useful to understand the characteristics of each group and could help design classifiers based on the gained knowledge. ### Limitations The obstacles in the way of ccPCA lie within it's limitations, a large one being time and space complexity for problems of growing size. Since we use cPCA, we will need to address its limita- tions in terms of time and space complexity for a large scale problem. In comparison, PCA uses **incremental algorithms** to solve this issue. ### Future Prospects In order to combat the problem above, there is admittedly a prospect to develop an incremental version of cPCA. ### Conclusions Dimensionality reduction is widely used to analyze high dimensional data for pattern discovery and real-world problem-solving. The cPCA clustering makes a tangible contribution to interpreting and understanding DR results by introducing a visual analytics method that capitalizes on contrastive learning. Using a scalable visualization, the method directs the user to the essential features within the data. cPCA, thus, further enhances the usability of DR methods. `39:[{"id":"db1736752a47654d280ed7f474af7e75df865cfeb8ddac355b08449af656893c181708d04ea8234385f8a54b5389a8dce6dc3d7c1dfb7ca479c3e21c29edc24a","href":"https://static.observableusercontent.com/files/db1736752a47654d280ed7f474af7e75df865cfeb8ddac355b08449af656893c181708d04ea8234385f8a54b5389a8dce6dc3d7c1dfb7ca479c3e21c29edc24a","name":"workflowEg4.png","type":"image/png","lastModified":1585977462702.762,"size":587554},{"id":"5863a6dbbfb1d1b50bf64a5e9d27ba66637ca97223165bd87ee4d55ffc82e0adc51e179f55187b59612cf2f0b0f037d4d8cd1321c28958702d4f1d7851704d79","href":"https://static.observableusercontent.com/files/5863a6dbbfb1d1b50bf64a5e9d27ba66637ca97223165bd87ee4d55ffc82e0adc51e179f55187b59612cf2f0b0f037d4d8cd1321c28958702d4f1d7851704d79","name":"network-vis.png","type":"image/png","lastModified":1585964017334.208,"size":948827},{"id":"1f629febd184e51a7833353f132f1bf013149b358d8f729d51a25fbdfb9a53543ddb837bbe52acb4613f6cc6014ed06b31d55766c3be4ad1128ecd8720a1911d","href":"https://static.observableusercontent.com/files/1f629febd184e51a7833353f132f1bf013149b358d8f729d51a25fbdfb9a53543ddb837bbe52acb4613f6cc6014ed06b31d55766c3be4ad1128ecd8720a1911d","name":"digits.png","type":"image/png","lastModified":1585968441331.569,"size":472873},{"id":"409588ecf8f5c2c10be70dd2c3433b99712b637703c2056cc602d6b6b164618c3b2e6fc0f8a1b4cb9af8f3688f044fbab7c7b1f0823c0257f361e8f795a41897","href":"https://static.observableusercontent.com/files/409588ecf8f5c2c10be70dd2c3433b99712b637703c2056cc602d6b6b164618c3b2e6fc0f8a1b4cb9af8f3688f044fbab7c7b1f0823c0257f361e8f795a41897","name":"parallelCoordinates.png","type":"image/png","lastModified":1585968540373.492,"size":1124490},{"id":"89ac0dc970674ee23ed1043adc1645891b270c6b00b55d79cc86b5b413dab9c212efe2c70df05844f0324dfbb4ef43950f01091ac5c691a93d36b46901709fed","href":"https://static.observableusercontent.com/files/89ac0dc970674ee23ed1043adc1645891b270c6b00b55d79cc86b5b413dab9c212efe2c70df05844f0324dfbb4ef43950f01091ac5c691a93d36b46901709fed","name":"workflow.png","type":"image/png","lastModified":1585968922546.946,"size":124084},{"id":"187213d8ab35a4c232745a312ad27ca9cf26a0d61166bdef81a54cfeba6720da14e91580de5c60f16f0011612d54bb187faa45f682762436574ea5157bf15be9","href":"https://static.observableusercontent.com/files/187213d8ab35a4c232745a312ad27ca9cf26a0d61166bdef81a54cfeba6720da14e91580de5c60f16f0011612d54bb187faa45f682762436574ea5157bf15be9","name":"workflowEg.png","type":"image/png","lastModified":1585970397683.955,"size":589008},{"id":"f647b761aa217041b35f05e844e64dcc2171a0a5d7cd05f6d202f50eeb68d761041982b5541e6a19863e7739d607d97791b8de3f743de1eef5b9e42c8959b0e3","href":"https://static.observableusercontent.com/files/f647b761aa217041b35f05e844e64dcc2171a0a5d7cd05f6d202f50eeb68d761041982b5541e6a19863e7739d607d97791b8de3f743de1eef5b9e42c8959b0e3","name":"workflowEg3.png","type":"image/png","lastModified":1585977462703.994,"size":591572},{"id":"f764004b26068f6840a71c2e471b209141354cd6c45cc2c749f04b675f0cfef32d3cdc86a6d6dac9ab117c0c6216b802609ad710c8f812a4043b4d19b5cee97f","href":"https://static.observableusercontent.com/files/f764004b26068f6840a71c2e471b209141354cd6c45cc2c749f04b675f0cfef32d3cdc86a6d6dac9ab117c0c6216b802609ad710c8f812a4043b4d19b5cee97f","name":"neural-network2.png","type":"image/png","lastModified":1585976540779.674,"size":459397},{"id":"98fb0837714d2c79cbc277dfb333f7a380708388bf14d4411983847b74975f34e4e17cf7a11ed84f314214fc1481f68d3fb931f76f1ecc68c85f9fdd4b3fe4b7","href":"https://static.observableusercontent.com/files/98fb0837714d2c79cbc277dfb333f7a380708388bf14d4411983847b74975f34e4e17cf7a11ed84f314214fc1481f68d3fb931f76f1ecc68c85f9fdd4b3fe4b7","name":"workflowEg2.png","type":"image/png","lastModified":1585977462508.66,"size":384521},{"id":"de7b34c3263d17dcfec7269e2ee0c0edd75582b6bae8f045b19ffc23b2ea09c7aba15ef06a57b7f4a1b7a896785accdc3316bcee0fb138c68a3ce02297289ec6","href":"https://static.observableusercontent.com/files/de7b34c3263d17dcfec7269e2ee0c0edd75582b6bae8f045b19ffc23b2ea09c7aba15ef06a57b7f4a1b7a896785accdc3316bcee0fb138c68a3ce02297289ec6","name":"neural-network.png","type":"image/png","lastModified":1585975450228.72,"size":350387},{"id":"4a36b7ea79cab655c96b81998af75e572a08d2dde29c358665a1a2ff846ec46faafea10c219e67d70fb46ce83abee828c9dd23b5040d46c9cdcbef9b2a6af1ae","href":"https://static.observableusercontent.com/files/4a36b7ea79cab655c96b81998af75e572a08d2dde29c358665a1a2ff846ec46faafea10c219e67d70fb46ce83abee828c9dd23b5040d46c9cdcbef9b2a6af1ae","name":"profile.jpeg","type":"image/jpeg","lastModified":1585975203293.514,"size":1742415}] 31:{"id":"4e4df706d8814cae","slug":"the-role-of-contrastive-learning-in-visualization","title":"The Role of Contrastive Learning in Visualization","thumbnail":"f89d248a44efceb0be51c23f4d65b6b172e44fa5fcd6cfc2e256fb84e05bb0d7","access_level":"public","create_time":"$D2020-04-02T19:10:23.504Z","update_time":"$D2020-04-04T05:54:44.995Z","view_time":null,"first_public_time":"$D2020-04-04T02:26:40.813Z","has_unpublished_edits":null,"fork":true,"stars":1,"starred":false,"owner":{"type":"team","id":"83b8aca8a4cdd0a3","login":"pkhambat1","name":"Pezanne Khambatta","avatar_url":"https://avatars-next.observableusercontent.com/80813e50fb491c13bf366ded74b6172351ad3eaf06e8002ab4d15e3b99727c23","tier":"starter_2024"},"version":463,"publish_version":463,"role":null,"authored":false,"authors":[{"type":"individual","id":"3d7cc0dad4047830","login":"pkhambat1","name":"Pezanne Khambatta","avatar_url":"https://avatars-next.observableusercontent.com/80813e50fb491c13bf366ded74b6172351ad3eaf06e8002ab4d15e3b99727c23","tier":null}],"members":[{"type":"individual","id":"3d7cc0dad4047830","login":"pkhambat1","name":"Pezanne Khambatta","avatar_url":"https://avatars-next.observableusercontent.com/80813e50fb491c13bf366ded74b6172351ad3eaf06e8002ab4d15e3b99727c23","tier":null,"role":"editor"}],"comments":[],"custom_thumbnail":false,"trashed":false,"lapsed":null,"subscribed":null,"forks":1,"collection":null,"type":"notebook","body":{"titling":"auto","title":"The Role of Contrastive Learning in Visualization","cells":[{"id":0,"value":"md`# The Role of Contrastive Learning in Visualization \n\nA Research Paper Summary & Critique `","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":281,"value":"md`\n\n

${Object.assign(await FileAttachment(\"neural-network.png\").image(), {style: 'width: 100%; max-width: 700px;' })}\n

Image Source: [https://www.kdnuggets.com/2018/02/8-neural-network-architectures-machine-learning-researchers-need-learn.html](https://www.kdnuggets.com/2018/02/8-neural-network-architectures-machine-learning-researchers-need-learn.html)

\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":309,"value":"md`\n**TLDR**: Critiquing the visualizations used in a research paper that talks about an effective method to cluster multi-dimenisonal data \n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":17,"value":"$33","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":338,"value":"$34","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":159,"value":"$35","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":374,"value":"$36","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":377,"value":"$37","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":20,"value":"$38","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":427,"value":"md`\n

${Object.assign(await FileAttachment(\"neural-network2.png\").image(), {style: 'width: 100%; max-width: 700px;' })}\n

Image Source: [https://www.pinterest.com/pin/47147127328202798/](https://www.pinterest.com/pin/47147127328202798/)

\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":350,"value":"md`\n## Acknowledgements\nFigures 1 - 7 are sourced from the original research paper.\n\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":44,"value":"md`## About the Author\n \n

\n\n\n👋 Hello!\n\nI'm Pezanne Khambatta, a student pursuing a Masters in Computer Science at [University of San Francisco](https://www.usfca.edu/). I graduate in 2022. Data Visualizaion has been a growing passion for me because it is so powerful in the way it transforms data into something intelligible by humans. Besides software development, I am passionate about good design (product and UX), music and photography.\n\n*As an aside, I've been obsessing over [algorithm visualizations](https://bost.ocks.org/mike/algorithms/) over the last few days.* \n\nI chose the topic of Dimensionality Reduction because it is a non-trivial problem for visualizers to choose which parts of a dataset to strip entirely or how to consolidate some parts, to promote better readablility.\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"}],"files":"$W39","stdlib":"1"}} 32:"$31:body"