alexbeloi

One way to optimize is to maximize expected revenue - expected cost of false positives or fraud.

A savvy business will know (or can estimate): customer lifetime value, false/true positive/negative rates of their fraud detection system, rate of charge-backs, expected rate of fraudulent purchases, revenue from given suspected transaction.

If average discounted customer lifetime value is $10k, charge-back rate is 2%, your fraud detection false positive rate is 0.1% and true negative rate is 99.9%, fraud detection true positive rate is 95% and false negative is 5%, customer is purchasing a $20 item. Then

* expected revenue if purchase is fraudulent: $0 * (true positive rate) - $20 * (false negative rate) = -$1

* expected revenue if purchase is non-fraudulent: $20 * (true negative rate) - $10k * (false positive rate) = $9.98

* total expected revenue value (with fraud detection enabled): (expected revenue if purchase is fraudulent) * (rate of fraudulent purchases) + (expected revenue if purchase is non-fraudulent) * (1 - rate of fraudulent purchases) = $9.7604

Without fraud detection, your expected revenue is: $20 * 0.98 = $19.6

Simplifying assumptions: false positive results in complete loss of customer value (realistically, replace this with big drop in customer lifetime value). Fraud rate is constant (realistically, should be modeled). Fraud rate is charge-back rate.

In this case, it's easy to see that seemingly low 0.1% false positive rate is still too high for this small of a purchase and these customer lifetime values. The 'smart' decision would be to ignore fraudulent purchases of this size in this case. (for this scenario, you need FPR below 0.004% with all else same)

Better model still would be a fraud detector that outputs a confidence score rather than "yes/no", and use the formula above to determine if the predicted false-positive-rate at this confidence level is sufficiently high to expect a revenue uplift from enabling the detector.

Using DUIs as an example, these cover things besides alcohol, can be prescription drugs, probably other things that impair motor function (IANAL).

Roadside DUI test is a general test for motor impairment. BAC tests are a nice proxy for that in the case of alcohol, and a legally convenient one as they are precise and objective, but the police care (morally speaking) about your motor impairment not your BAC.

tl;dr, we can already answer the question "is this person too stoned to drive?", roadside DUI test.

>> especially newer coders, women, people of color, and others in marginalized groups.

>But this isn't true.

What do you make of their statement that "Many people, especially those in marginalized groups do feel less welcome. We know because they tell us."?

>I respectfully argue that until you can make Diesel as clean as H2 fuel cell, the push to eliminate diesel burning engines will go on.

What market are these competing in?

Graduate students unionizing is the norm in public universities. As the article points out, it's only recently (2016) that the NLRB ruled that graduate students at private universities are to be considered employees, which opened the door for students to vote whether they wanted to unionize, prior to this they had no legal standing to even consider unionization.

[0]https://en.wikipedia.org/wiki/Graduate_student_employee_unio...

Who is your target market?

Why does a payment gateway need deep learning? I can't imagine an online betting system doing fraud detection on deposits.

> how much do the "obvious" checks help - highly unlikely purchases (3 iphones) timing or physical activity (I probably won't buy books on amazon, clothes on boohoo and petrol in a garage in the same five minutes) versus the more ML / secret squirrel stuff?

ML solutions very often are just learning to codify these 'obvious' scenarios, and as a bonus sometimes less obvious ones.

You could sit in meetings for hours/days listing all the cases the fraud detection should catch, and you'll still end up missing lots. If you're stripe, you have tons of data about fraudulent purchases that you can use to learn and codify these scenarios. Importantly, you can learn which scenarios are most prevalent in your system in particular, fraud at Stripe is very likely to look different than fraud at (say) Wells Fargo.

I agree that this is a bad way of measuring humans accurately, but I would say that _accuracy_ is not exactly the correct or optimal goal for hiring.

Google (and other's) strategy is to optimize for precision rather than recall or accuracy, which is the optimal strategy when you receive several orders more candidates than you are capable or willing to hire.

To maximize precision, you need to minimize Type 1 errors (False Positives), and you don't really care about Type 2 errors at all (False Negatives, e.g. throwing the baby out with the bath water).

This is of course if you treat people like data points, which seems more and more common (and explicit) with larger companies.

Yeah... with 1000 conflicting versions of any event, it's more difficult than ever to sift through the misinformation. Then one brilliant soul will say "aha, I know the solution, I will hire experts and engineers to efficiently sift through information sources and present people with a version of truth we can be confident about!", behold the birth of the 1001st version of events.

As the author points out in the conclusion, the state space blows up very quickly as the grid becomes larger.

There is a large class of algorithms for finding approximately optimal solutions to MDPs[1] that are model-free or stateless, meaning you don't need to enumerate all of the state-to-state transitions to get a good policy.

If you google 2048 reinforcement learning[0], you'll find lots of implementations of such algorithms.

[0] https://www.google.com/search?q=2048+reinforcement+learning

[1] https://en.wikipedia.org/wiki/Markov_decision_process#Algori...

>its CEO has strong personal reasons for valuing [privacy]

I don't know what this is referring to, could you please share?

I agree, your sentiment is precisely what I was implying with the quotes.

These things don't always communicate perfectly through text or punctuation.

Or someone that's "just doing their job".

They used it to predict political affiliation of people that don't explicitly state a party preference.

The two parties already have a list of registered party members (and they can see who on Facebook explicitly states their party preference), for those members the main goal is higher turnout (they are the training data). The other voters they're interested in are unregistered (e.g. independent) voters that are likely to be on their side ideologically.

The core idea is very simple, they believe that if someone says they're independent, but their preferences/features (age, gender, location, likes, posts) are predicting moderate or high likelihood of $PARTY affiliation, then showing this person political ads may move them from 'maybe vote for $PARTY' category and get them in the 'definitely vote for $PARTY' category.

If you have continuous access to new Facebook data as you're serving ads, you can verify your ads are working on an individual basis by checking the predicted 'score' for $PARTY affiliation predicted by your model before and after an ad (I want to stress that this can be done on an _individual basis_). The likely sequence of events is that they did AB testing on different kinds of ads and found that fake inflammatory ads were most effective at achieving this goal in a very measurable way ($PARTY score), the resulting media/political atmosphere is collateral damage (hopefully unintended).

Source: I am a data scientist / machine learning scientist and this is how I would do it and how it seems to me others. I don't work on political data but I have worked on personalized recommendations which are similar.

I agree with you (regarding gun owners as a protected class) and I believe our two claims are complementary rather than adversarial.

Unlike victims of gun crimes, the victims of copyright violation have the money and lobbying power to effect change.

>Wasabi and B2 are pretty much the only commercial services which could even attempt to compete on price with something like Sia today (Filecoin and others in the future potentially)

It would be nice to see Sia/Filecoin provide some basic DB metrics like availability, durability, latency, bandwidth... things that any prospective client would absolutely need to know.

I'm no expert so bear with me.

My understanding is that the danger of miner control is not "double-spending" but making multiple spends and being able to pick the most favorable one after the fact.

As a non-controller if you try to double-spend, you can't know which one of your attempted transactions will be recorded in the ledger and which will be rejected. As a 51% controlling miner, you can choose the 'correct' transaction.

This allows the 51% controlling miner to engage in futures arbitrage.

Could you provide a couple sources please, I believe many of us are not well read on blockchain literature.

I'm not super familiar with ipfs, is that sufficient to cover yourself legally speaking? Does the company host any server nodes itself?

You're using "open" in different contexts, one applying to a platform (e.g. blockchain or facebook app api) and another in regards to access control.

As platforms, blockchain is more open than facebook and for access control it's less open. GP was referring to security being at odds with openness in the sense of access control.

>Always have a naive solution to compare against the AI solution. 'Naive' here may be a simple linear regression, RMSE, or multi armed bandit/Thompson sampling.

This cannot be stressed enough, optimism bias will always push the scientist towards the 'more interesting/complete/new' method and model, but a seasoned practitioner will have the discipline to always establish a baseline (<1 days work).

Estimated 57,700 slaves in the United States according to Global Slavery Index [0]. Here is a related WP article[1].

[0] https://www.globalslaveryindex.org/index/

[1] https://www.washingtonpost.com/news/worldviews/wp/2013/10/17...

>It seems obvious to me what really needs to be fixed is a broken electoral process.

Could you propose an alternative that addresses your points.

It's the same as the difference between societal discrimination and government legislated discrimination. Both terrible, the latter seems more so.

Yes, that is the core of value based pricing: https://en.wikipedia.org/wiki/Value-based_pricing

In principal Bart[0] is supposed to be operated almost automatically and the train operator is there for announcements, closing doors and watching the track for hazards. In practice, the operator does a lot of speeding up/down manually. I can't find it now, but I remember reading an article that claimed that Bart trains were operated in manual mode 90% of the time, not sure the reason.

[0] https://en.wikipedia.org/wiki/Bay_Area_Rapid_Transit#Automat...

Paypal represents 21% of e-commerce transactions in the UK and 40% in Austria, similarly popular in many other European countries where debit/credit is widely accessible. I wouldn't call that niche.

[0]https://ecommercenews.eu/ecommerce-per-country/ecommerce-the... [1]https://www.statista.com/statistics/434121/online-shops-popu...

Is brand partner another word for advertiser/marketer?

I see a couple image ads on the reviews, but I doubt that brings in much revenue. So are doing things like product placement/mentions in-article?

I hope you don't accept money from anybody pushing you to review particular restaurants.