{"id":25447,"date":"2020-11-12T00:00:00","date_gmt":"2024-09-13T08:48:45","guid":{"rendered":"https:\/\/wordpress-583806-4798031.cloudwaysapps.com\/7-red-flags-in-another-data-misinterpretation-wave-will-ai-mitigate-it\/"},"modified":"2024-11-15T16:21:40","modified_gmt":"2024-11-15T15:21:40","slug":"7-red-flags-in-another-data-misinterpretation-wave-will-ai-mitigate-it","status":"publish","type":"post","link":"https:\/\/informator.se\/en\/7-red-flags-in-another-data-misinterpretation-wave-will-ai-mitigate-it\/","title":{"rendered":"7 Red flags, in Another Data-Misinterpretation Wave – Will AI Mitigate it?"},"content":{"rendered":"\n

Statistics climbed from an unpopular school subject to top of the agenda in 2020, and skilled data analysts are in short supply in the year of astoundingly-exponential curves. Although health<\/a>\u2013related AI and data analysis expand, many surprisingly well-informed academic graduates still get cheated by rush-job presentations of correct data compiled from reliable sources. This incomplete list of red flags provides some inspiration for your own rules of thumb and your news assessment.   <\/em><\/p>\n\n\n\n

1. Curves on \u201cfavorably incomplete\u201d scales<\/strong><\/h3>\n\n\n\n

– sometimes to save space on a page, but sometimes it seems intentional:<\/p>\n\n\n\n

\"redflag\"<\/figure>\n\n\n\n

2. A rush to publish absolute figures (and sometimes even compare them\u2026)<\/strong><\/h3>\n\n\n\n

This is becoming frequent in countries that stand out in Covid-related fatalities per100,000. Even if correct, absolute figures aren\u2019t comparable; sometimes they rather conceal than inform.<\/p>\n\n\n\n

\"redflag\"<\/figure>\n\n\n\n

In Sweden: extend the left bar 138-fold<\/strong> first, to make the two comparable, before drawing any conclusions. (On Iceland: 3,940<\/strong>-fold. Etc.) Or, simply go to reliable<\/a> sites that offer sorting of the tables by a population-adjusted <\/strong>column.<\/p>\n\n\n\n

3. \u201cExponential\u201d as a synonym of steep.<\/strong><\/h3>\n\n\n\n

Even linear growth can be steep, which is easy to see, although its pace is constant:<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Exponential growth requires more thought and imagination: it doesn\u2019t start so steep, and given an exponent close to 1 it doesn\u2019t seem steep at all:<\/p>\n\n\n\n

\"redflag\"<\/figure>\n\n\n\n

However, suppose we\u2019ve only ten infected patients, a modest R value of 1.15<\/strong> , and we estimate their subsequent cases, via 1st<\/sup>-hand infection and way \u201cdown the road\u201d to 10th<\/sup>-hand. The sum total along these ten trees (pathways) will exceed 200<\/strong>. But, assuming a slightly increased R of 1.25 instead, it\u2019ll exceed 330. Unlike a steep linear slope, exponentials cheat our intuition, and cheat decision makers into delayed inconsistent lax decisions.<\/p>\n\n\n\n

Moreover, Corona viruses are a moving target, because they mutate<\/a> fast. Distancing, discipline, strict steps and measures \u201ctame\u201d them in a matter of weeks, and gradually reduce lethal cases to near zero. The reverse, delayed and lax measures increase the R and genetically \u201cfavor\u201d aggressive strains instead\u2026<\/p>\n\n\n\n

4. Mistaking a correlation for causality<\/strong><\/h3>\n\n\n\n

It\u2019s raining when gardeners turn their sprinklers off. True, but figuring out the direction<\/strong> of the cause-effect relationship takes some experience. Kids tackle it by learning and common sense. AI R&D is tackling it by several methods, e.g. logical exclusion (ruling out, for example, a common third cause, causality chains, etc), and by widening it from Narrow AI toward Full AI and commonsense reasoning. Using Wikipedia and the Internet as a vast training-data lake takes us to (many) baffling trifles of<\/strong> Big Data<\/strong>.<\/p>\n\n\n\n

5. Leveling up<\/strong><\/h3>\n\n\n\n

Donald Michie<\/a>, meant that information mining uncovered the tacit know-why hidden in<\/strong> human know-how<\/strong> (represented in statistics about previous decisions). In its training phase, the ML algorithm became a hypothesis creator. Then, during test and field test, the ML-generated logic became a hypothesis tester.<\/p>\n\n\n\n

However, as the training datasets scaled up from thousands of rows to millions or billions (and the number of columns, i.e. attributes, expanded as well), the number of possible correlations <\/strong>exploded; some of them were meaningful, others extremely far-fetched. This was often nicked \u201cmass significance\u201d (so, disease gravity could correlate with trifles like, say, shoelace color on patients\u2019 shoes).<\/p>\n\n\n\n

6. Input:  mass<\/strong> versus relevance and accuracy<\/strong><\/h3>\n\n\n\n

Suppose in targeted<\/strong> Covid-testing, we focus on high-risk segments such as paramedics, doctors and nurses. Now assume that, regrettably, 96% of these are really<\/strong> infected. Then, most of Covid-positive <\/strong>test results are real <\/strong>positives, and only a couple of percent are false positives<\/strong>, because of the extreme situation in those well-chosen segments of our real world: <\/p>\n\n\n\n

\"redflag\"<\/figure>\n\n\n\n

Now, more time-consuming lab tests double-check and confirm that our test-campaign figures were quite accurate. So we scale it up, right away, and test everybody in our country. Now assume that, luckily, 96% of our 20 million-population is negative (not <\/strong>infected). Hence most of Covid-negative <\/strong>test resultsare real <\/strong>negatives, and only a couple of percent are false negatives<\/strong>, because of the fortunate situation out there. Again, lab tests confirm it\u2019s correct.<\/p>\n\n\n\n

\"redflag\"<\/figure>\n\n\n\n

Too good to be wholly true? Well, in the 1000-times bigger data (20 million tests), those red triffle-percent false negatives<\/strong> at the top (that is, of Covid-infected \u201cmisses\u201d, still spreading<\/strong> the virus) can approach half a million<\/strong>, although most positives <\/strong>were false alarms now (again because of the auspicious situation in the country as a whole). So, the colors from the first round (the risk-group tests) are now reversed on the positives side.<\/p>\n\n\n\n

\"redflag\"<\/figure>\n\n\n\n

Consequently, even if the number of quarantined people might happen to be approximately right, most of them are nevertheless the wrong guys. Whereas, up to half a million of infected false negatives are running around…<\/p>\n\n\n\n

7. A level-up: The fuzziness of real-world variables<\/strong><\/h3>\n\n\n\n

The scaled-up part (of 6. above) raises a question: how much<\/strong> of a real danger <\/strong>are all those false negatives out there?<\/p>\n\n\n\n

Our data analyst points out the objectives of our test campaigns: In the small segment, it was primarily correct diagnosis <\/strong>and treatment<\/strong> of infected staff, and we could even afford some pre-test diagnostics and repeated Covid-tests minimizing false negatives further, to get a clear No or Yes (sick) for each individual. On the other hand, the objective of the scaled-up tests is primarily tracking<\/strong>, to isolate outbreaks; this calls for attention to a grey zone<\/strong> between Yes and No: people with virus particles per million (ppm) near limit<\/strong> (for Covid: 100,000 ppm), where the risk of transmission <\/strong>nears zero.<\/p>\n\n\n\n

Time-comsuming accurate tests are very sensitive (down to a couple of hundred ppm), which was important in the diagnostics, but now we\u2019re fine with down to 100,000 for tracking purposes; here, paradoxically, not-so-good sensitivity of quick tests<\/a> comes in useful (makes this bug a feature, so to say) , because positives that are no longer infectious are less interesting to track. This narrows our grey zone, although not erases (it\u2019s still a zone, not a hard fine line). We have to live with some fudge-factor in a real world where 99,999 ppm doesn\u2019t mean 0.0% risk of transmission, and 100,001 doesn\u2019t mean 100.0% risk. So, our recent \u201cbinary\u201d classification (cases either Yes or No) is becoming akin a fuzzy truth function.<\/p>\n\n\n\n

\"redflag\"<\/figure>\n\n\n\n

Last: remember the two-way street between ML and data<\/strong><\/h3>\n\n\n\n

Each level-up of complexity makes our brains a bit more dizzy. My point is: unlike our AI systems.<\/p>\n\n\n\n

ML <\/strong>not only constrains, <\/strong>but also assists <\/strong>data architects. Therefore, \u201cantipatterns\u201d similar to the first six red flags are something AI systems can help us recognize, pattern-match to recent cases, and warn for (in data presentation and elsewhere). They can also provide useful advice as complexity grows (like 7. above, for example). Likely, similar pattern recognition & matching will become part of data pipelines<\/a> and DaaS\/MLaaS, along with pattern-matching inventions<\/a> within hard-core database-tech for Big Data (e.g. \u201clearned indexes\u201d by Google, for fast lookups). The cost of the training phase in ML is only a fraction of the pipeline sum total. Think: data ingestion, storage, preparation, ML, deployment, field test, governance, upgrades\u2026<\/p>\n\n\n\n

Therefore, quite some automation of data pipelines will augment ML and data architecture for years to come. Innovation sparked by Covid-research will turn out helpful in many other kinds of monitoring, forecasts, prevention, emergencies, and industry sectors.<\/p>\n\n\n\n

Stay safe!<\/p>\n\n\n\n

by Milan Kratochvil<\/a><\/em><\/p>\n\n\n\n

<\/a> Milan’s courses:<\/strong><\/p>\n\n\n\n

AI, Architecture, and Machine Learning<\/a><\/em> (November 2020)<\/em><\/p>\n\n\n\n

Agile Architechture Fundamentals<\/a><\/em> (December)<\/p>\n\n\n\n

Agile Modeling with UML<\/a><\/em><\/em><\/p>\n\n\n\n

(on demand: Modular Product Line Architecture<\/em> )<\/p>\n\n\n\n

(on demand: Avancerad objektmodellering med UML<\/a><\/em> )<\/p>\n<\/span><\/span>\r\n

\u00a0<\/span><\/span><\/h4>\r\n<\/span><\/span>","protected":false},"excerpt":{"rendered":"

Statistics climbed from an unpopular school subject to top of the agenda in 2020, and skilled data analysts are in short supply in the year of astoundingly-exponential curves. Although health\u2013related AI and data analysis expand, many surprisingly well-informed academic graduates still get cheated by rush-job presentations of correct data compiled from reliable sources. This incomplete […]<\/p>\n","protected":false},"author":6,"featured_media":26528,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_monsterinsights_skip_tracking":false,"footnotes":""},"categories":[571,570,922],"tags":[634,592],"class_list":["post-25447","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-blogg","category-blogi","tag-big-data","tag-machine-learning"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t