I just used this at work the other day to calculate similarities between different data models that had overlapping children models. One of our teams was going to go through manually to check these overlaps and consolidate, but by using this clustering algo based on Jaccard distance we were able to give them clusters to consolidate up front. Super cool stuff!
Jaccard Index
11–20 of 46 posts
Re: Jaccard Index
#12What is the predicted bounding and the ground truth bounding, as related by a stop sign? I have no idea what's happening there.
Re: Jaccard Index
#13I just used this at work the other day to calculate similarities between different data models that had overlapping children models. One of our teams was going to go through manually to check these overlaps and consolidate, but by using this clustering algo based on Jaccard distance we were able to give them clusters to consolidate up front. Super cool stuff!
what is a children model? I'm curious but can't really follow what you wrote, can you add a bit more context?
Re: Jaccard Index
#14I recently wrote a fun blog post (https://pncnmnp.github.io/blogs/odd-sketches.html) about how to estimate Jaccard Similarity using min hashing, what b-bit min hashing is, and how to improve upon its limitations using a 2014 data structure called odd sketches.
Jaccard Similarity's history is also quite interesting. From my blog:
> In the late 19th century, the United States and several European nations were focused on developing strategies for weather forecasting, particularly for storm warnings. In 1884, Sergeant John Finley of the U.S. Army Signal Corps conducted experiments aimed at creating a tornado forecasting program for 18 regions in the United States east of the Rockies. To the surprise of many, Finley claimed his programs were 95.6% to 98.6% accurate, with some areas even achieving a 100% accuracy rate. Upon publishing his findings, Finley's methods were criticized by contemporaries who pointed out flaws in his verification strategies and proposed their solutions. This sparked a renewed interest in weather prediction, which is now referred to as the "Finley Affair."
> One of these contemporaries was Grove Karl Gilbert. Just two months after Finley's publication, Gilbert pointed out that, based on Finley's strategy, a 98.2% accuracy rate could be achieved simply by forecasting no tornado warning. Gilbert then introduced an alternative strategy, which is now known as Jaccard Similarity.
> So why is it named Jaccard Similarity? As it turns out, nearly three decades after Sergeant John Finley's tornado forecasting program in the 1880s, Paul Jaccard independently developed the same concept while studying the distribution of alpine flora.
Re: Jaccard Index
#15What is the predicted bounding and the ground truth bounding, as related by a stop sign? I have no idea what's happening there.
The "predicted" box there would be a best guess from a statistical model powered by AI or computer vision, answering, "where is the stop sign in this image?". The "ground truth" would be an annotation by a human answering the same question. The jaccard similarity metric would say that these bounding boxes are highly similar, and so the prediction could be evaluated as high quality.
Re: Jaccard Index
#16The name may be an example of this: https://en.m.wikipedia.org/wiki/Stigler%27s_law_of_eponymy It was developed by Grove Karl Gilbert in 1884 as his ratio of verification (v)[1] and now is frequently referred to as the Critical Success Index in meteorology.[2] It was later developed independently by Paul Jaccard…
Another example of this sort of thing (that is vaguely related in that it's commonly used as a metric) is (what I call) the Matthew's Correlation Coefficient https://en.wikipedia.org/wiki/Phi_coefficient
> In machine learning, it is known as the Matthews correlation coefficient (MCC) ... introduced by biochemist Brian W. Matthews in 1975.[1] Introduced by Karl Pearson,[2] and also known as the Yule phi coefficient from its introduction by Udny Yule in 1912
Re: Jaccard Index
#17I recently used Jaccard similarity as a measurement of distance between two sets of online articles. It’s amazing how versatile it is for all sorts of weird tasks.
I uses to use Jaccard similarity combined with w-shingling at the character level to detect clusters of fraud sites. It was surprisingly effective, because it was able to pick up common patterns in the code even if they used completely different styles and text. https://en.m.wikipedia.org/wiki/W-shingling
Interesting - I also used Jaccard similarity to classify clusters of malicious ad traffic schemes. The idea worked well. It was unclear if the similarity was due to mimicry or authorship, but that did not matter for our use.
Re: Jaccard Index
#18I think a good example of the Jaccard Index is from the paper Lazy Prices[1] in which the authors use it to measure diffs in a company's annual and quarterly reports. They also use other similarity measures (cosine, minimum edit, simple) so you get a better understanding of how each applies to textual similarity as a whole.
[1] https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1658471
Re: Jaccard Index
#19I've used this to find similarity between terms in a specification document and identifiers in a program. The index is computed on the bag of bigrams from the two strings.
If one has a set of pairs that are similar, one can look for common bag differences in the matches. These can correspond to extra characters inserted in the identifier names for that particular program (for example, prefixes or suffixes related to module name or variable types.) Once these are found they can be used to tweak the similarity score.
Re: Jaccard Index
#20I recently wrote a fun blog post ( https://pncnmnp.github.io/blogs/odd-sketches.html ) about how to estimate Jaccard Similarity using min hashing, what b-bit min hashing is, and how to improve upon its limitations using a 2014 data structure called odd sketches. Jaccard Similarity's history is also quite interesting. From my blog: > In the late 19th century, the United States and several European nations were focused…
Recently I wanted to try and use this to filter out logs I don't care about, but it seemed a lot more involved than I initially thought.
I essentially wanted to use this as a way to flexibly filter out items without having to come up with a regex for every line item.
I wonder if anyone has done this before...