A case where Germany has sold crapware to the Swiss. I mean geographic analysis of crime can certainly be helpful, but I doubt it can be more accurate than your average fortune teller if the rate of crime is as low as it is.
The the software is named precobs in reference to minority reports precogs let's me believe their marketing team is hired from people selling penis enlargement on the internet. Perhaps they are just smarter because they know how often people are inclined to buy bullshit.
A doctor once said in some interview that if there really was a pill or treatment to enlarge your penis, you wouldn't need to market it through email, because everybody would know about it. I think crime prediction would follow the same rule.
Exactly. If the thing you are trying to predict is very rare (compared to business as usual), the error rates of your system have to be unfathomably low if you don't want to either drown in false positives or detect nothing at all ( =expensive money wasting project).
In some cases the error rates need to be so low, that even claims like "we are 99.9% accurate" should be taken with a big crystalline block of salt.
Remember this whenever terrorism is used as a reason why we desparetely need these surveillance mechanisms. If burglars are rare, terrorists are even rarer. The proposed systems are good for bigger things like demonstrations, social movements etc. but they suck at detecting terrorism, unless you have a metric ton of personal which will look at false positives all day long.
And yet this reasoning is exactly why the police keeps reiterating that they need more and more power, both surveillance, and the power to interfere without a criminal cause.
Ironically this is actually true. The question is to what level you want crime to fall, and if the answer is zero, then you best let the police shoot every last person without requiring them to have committed a crime.
I remember when I lived in London there was a time when it seemed like there were more and more police sirens everywhere, this made me feel like there had been a big increase in bad people doing bad stuff.
But, it turns out, what had actually happened was they had changed the standard for sirens on emergency vehicles. As vehicles were modernised they all started to sound like police cars. At the same time a new system for medical emergencies was introduced that involved a first response by motorbike paramedic followed by backup a bit later from an ambulance if necessary, again all with sirens blaring. Suddenly it sounded like the police where rushing everywhere all the time. Whereas before you could tell the difference between a fire engine, an ambulance and a police car from the sound so you could tell what kind of incident was going on.
I'm sure that contributed to a sense that there was a big crime problem in the 90's when in actual fact it was starting to tail off.
Crime rate have very little to do with how society is policed. If you want to find explanations look at wealth distribution, poverty rates, the school system, job market and so on.
There was no statistically significant difference between Swiss regions that did use the predictive software and those that didn't. All regions saw a broadly equivalent drop in crime.
This is an education failure on the part of the purchasers. They should know what you can and cannot do with data. Think about the dataset you'd need to make the burglary prediction. Cops would have some priors about where/when it might happen, but is the data actually collected?
You'd need to know a lot of things about a large number of homes in order to see how the characteristics are distributed. Has the house been left empty for a while? Has someone posted about their holiday on FB? Is there something of value in the house? What security measures are there? Are known house-burglary gangs in action in the area?
Those are the kind of things you'd think cops might have priors on. I'm sure one can chime in on this. But they're also the kinds of things that you might have problems collecting data for. And even if you had it, your false positive rate needs to be pushed pretty low for it to be useful. Plus your false negative (place is burgled but system didn't say) needs to be really low, in a country where there isn't that much crime to begin with.
For the second system, as the article says, you can just say everyone is a risk. It's like saying everyone has Covid, you'll correctly find all the carriers. This is an obvious issue with ML systems that a purchaser needs to understand.
The last system, with the recidivism, is maybe the most concerning. It seems to suggest that people's paroles are decided based on this system? Then you'd better have good evidence that the system works, in the sensitivity/specificity sense.
This is the same as any other (ab)use of statistics. Forecasting is a valid and useful practice for managing your people and process. Maybe you redeploy resources to improve response time or schedule shifts to match forecasted incidents, etc. Data collection is pretty awful with this type of data, so good luck doing much more.
But the problem is that the public hears something different when they hear "automated crime prediction". People hear something novel that could prevent crime. After all, if rain is predicted, you grab an umbrella.
The reality is that cops on the street already know where crimes are likely to happen. You'll get some tools to reduce workload on police sergeants doing scheduling and things like data presentation are probably the primary benefit. Not as sexy as minority report.
> The reality is that cops on the street already know where crimes are likely to happen.
As for many jobs. I work as a consultant and I'm surprised how little effort companies do have conversations with their workers. People in the operational side of things is worth listening.
In fact, If I'm valued and I have clients that called me several times it's because I like to interviews, focus groups and informal conversations.
In the end, if I have to summarize my job, I listen to people and then do some stupid stuff with numbers to look professional and make my client swallow it. That's about it. It's 30k, thanks.
you could simply write a geographic correlation function on income/wealth and race/ethnicity and do just as well at this kind of forecasting, since police find crime where they look, and they tend to look in poorer and non-white areas.
further, we need to disabuse ourselves of the idea that we can predict crimes to any degree beyond random (vs. aggregate stats, like crime rates), because it's tantalizing ideas like that which leads to the slide into surveillance and totalitarianism without any real, positive return to society as a whole.
I don't believe anyone on HN is so naïve that they think that the main reason crime rates vary between areas is that police presence varies. Your post makes it sound like you are, but I assume that you only think varying presence skews the numbers rather than generates them?
i would hope no adult is naïve enough to believe police magically appear only where crime is and only addresses clear and existing crime, rather than being embedded in a complicated sociopolitical system overwhelmed by legal pitfalls. those are stories we tell children to comfort them for sleep.
> HN seems to think police spend their shift waiting for crimes to happen in front of them.
No, but when they aren't responding to calls, they spend the rest of their shift going on fishing expeditions.
What do you think traffic cops, for instance, are doing most of their day? What about the cruiser slowly driving through a neighbourhood? They aren't there to respond to a call, they are just running license plates/looking for poorly maintained cars to pull over (Obviously, if someone egregiously breaks traffic rules in front of them, they don't need to look very far - but when nobody's breaking the rules, they can always find someone who looks suspicious.)
The thing is, when you only go on fishing expeditions in poorer neighbourhoods/targeting poorer vehicles, you're going to find that will skew crime statistics - and statistical model based on that will direct even more policing towards those areas.
It's why in the United States, the higher you are on the income ladder, the more illegal drugs you consume - yet the lower you are on the income ladder, the more likely you will end up in prison for drug use.
that's a misdirection. most crimes are not emergencies, and policing encompasses much more than emergency response.
if policing were just, we'd see many more white-collar crimes like corruption, bribery, embezzlement, fraud, extortion, etc. being busted, because those crimes have much larger impact and wider fallout on communities and wider societies.
White collar crime isn't a task for police any more than tax evasion is - those just plain aren't tasks they are meant for as opposed to public safety tasks. Which also explains why they are less funded, even without any conflicts of interest they aren't "sexy" crimes that grab attention, they all also lack urgency and thus are easily "procrastinated". Digging through records can uncover many of them and there are a lot of them.
The police in their current role only show up at the behest of warrants in those situations to search or arrest.
Enforcement has an impact that can be measured for certain activities, but not all.
For example, enforcement of DWI laws had a meaningful impact on alcohol-related traffic fatalities. Deployments in NYC targeting crime of opportunity for things like muggings had a meaningful impact.
In some cases, there may be a halo effect where addressing certain problems may reduce others. But it's not a magic bullet. Drugs seem to be an example where enforcement has limited impact on outcomes. You also have the risk of heavy routine enforcement fostering corruption. A good example is the Border Patrol, which arrests ~250 officers annually.
> since police find crime where they look, and they tend to look in poorer and non-white areas.
That's true for property crime, moving violations, etc. But most of these crime predictors are looking at break ins, violent crime, murders, etc. Those numbers don't go up when police are focusing on an area, those generally go down.
> It tries to predicts burglaries from past data, based on the assumption that burglars often operate in small areas. If a cluster of burglaries is detected in a neighborhood, the police should patrol that neighborhood more often to put an end to it, the theory goes.
A review of 3 automated systems in use by the Swiss police and judiciary reveals serious issues. Real-world effects are impossible to assess due to a lack of transparency.
I guess it depends on what we would want from crime predictions.
Do we want a Minority Report type system where someone shows up right before the crime and prevents it?
They mention burglary. I'm guessing most buglers aren't seasoned professional buglers and rather it is a crime of opportunity. So if a cop is on the street they won't do it ... but maybe they'll do it the next street over, 20 minutes later?
Even if somehow there was better prediction, there really aren't enough police to patrol the target street, and the next one over, and the next one ... predictions may suppose infinite police resources would prevent it, but that's not a thing.
This might be a topic where 'prevention' isn't really possible.
I don't have numbers, but I don't think there is much home burglary as a crime of opportunity in Switzerland. The places worth entering, usually have good enough security measures, that you need to come prepared and it is going to take you some time. What you usually hear is certain neighboorhoods being scouted for lights being off for longer time, indicating the people are on vaccation. But I don't know in what regards crime predictions systems are useful for that, what value they add.
Repeat rates are VERY high in burglary. And a person who hasn't burgled is relative unlikely to commit a home burglary even if the opportunity presents. Some neighborhoods without repeaters here all leave their doors unlocked. I grew up in an area like this, never had a house key until I moved away, we would go on months long vacations without locking up the house. Literally EVERY neighbor could have walked in anytime (house was totally dark). It never happened.
Car thefts as well, I lived in a tough neighborhood, same guys checking cars all the time. If these guys were around just roll down windows and leave car unlocked. Still had my car stolen which was annoying.
In Germany it's very much not like that, organized groups (often coming from and returning to eastern european countries) make up a big chunk of burglaries and they are careful to choose areas with a high effort/reward ratio, which are rarely the low income areas.
I think because of low poverty and expansive social nets, crimes out of desperation are rare. E.g. Reducing crime is explicitly one of the main reasons why Switzerland mantains its heroin programme, providing the drug for free to people who have failed mutliple withdrawal attempts.
The risk/reward for small crimes just doesn't check out for almost everyyone. Why a lot of those is in the hands of organised crime that can scale and optimize it and pick out the people willing to do the "last mile" dirty work.
You might be confusing petty theft (pickpocketing etc.) with the legal of definition of burglary. Burglary involves breaking into property and then stealing something.
I don't know how you could possibly commit burglary as a crime of opportunity.
TBF this is because there is no crime in Switzerland.
The guys I work for are based in Zug ("Crypto Valley"). One time I was there, I went to a supermarket to buy some beers in the evening. At the self-checkout machine, as I was trying to pay, the machine asked me in German, "how old are you", with three options, 16-, 17, 18+. The 16- and 17 was in red and the 18+ was in green, and you just had to press the green button. No ID check or staff member coming around to look at you or anything.
Predicting crime is great. It means you're working to understand the state of your society.
Using it for policing is idiotic. You can't arrest a statistic.
Instead, it should be used for public policy, to understand what communities are at risk, direct efforts to understand why they are risk, and engineer better systems to support those communities.
There simply isn’t enough data to base predictions on. It will be close to random wether or not this system will “predict” a crime, to a similar degree that a clairvoyant can guess where the next crime will happen.
It seems like the law of small numbers at play here - or a snake-oil salesman doing a thorough job.
I tell people, the main use case for ML is to obfuscate and launder decision authority and accountability away from people and up into abstract organizations where nobody is accountable. If there is ever a question as to why the police made a decision, they can say it was their prediction model, which is so complex nobody could reasonably expect to understand it or be responsible for its outcomes.
It's as though artificial intelligence itself were a cryptological problem where it's only real when it becomes sufficiently complex that information about who is accountable for it is destroyed.
I think one of the most important things we can do is to establish liability for prediction models to the people using the model.
At the end of the day if you falsely arrest someone, fire them on bad grounds, etc then you're the responsible party. The excuse "Well but someone told me to!" doesn't work any better when you're blaming a machine instead of a person - it's still bunk.
If the model is broken, well then that's on you. A vendor sold you a magical model saying it was perfect and you got in trouble anyways? Well, you can go on and try to sue the vendor and hash that out - after you pay for your mistake.
It's critically important we don't allow models to become get out of jail free cards.
You can still go after a seller for making false claims, but you are liable for your actions - full stop. Unless the seller forced you to fire someone or arrest someone based on the model, you executed the action.
Someone who writes your legal contract isn't liable if you badly enforce it.
Discussions around responsibly applying probabilistic models to decision making at a societal level are really important to have.
Myth-making about ML as a shadowy conspiracy—instead of just another tool for engineering, like databases—is counter-productive.
The "main use case" for ML is the literal one, "designing models that can learn to predict high probability outcomes in situations where a deterministic model cannot be applied." It is another technology in an engineer's toolkit. Look at your smartphone, and you will see that the majority of apps you use rely on machine learning for ETA prediction, content recommendation, speech-to-text, image processing, translation, spam filtering, etc.
I think this is more government departments are easy to get money out of by convincing them your fancy computer words are better than their Excel spreadsheet.
The swiss had something like 50 murders TOTAL of which only 3 were unsolved. That's something like 0.5 per 100,000?
The USA is probably an ORDER OF MAGNITUDE higher?
The solve rates for crimes in the USA is horrendous - something like 35% in places like Baltimore.
Before we talk about how little the swiss have to show for their approach, perhaps we should allow them a touch of credit for creating a system that has reduced homicides AND led to the identification and conviction of those who commit them?
And maybe do something about the 15,000+ people killed per year in the US?
But it misses the larger context which is that Swiss are proactive in trying to reduce crime through many approaches. They use many automated tools, hotspot mapping etc. The article even mentions they use something like 20 tools.
As a result of this larger effort, they have kept crime low and solve lots of serious crime.
The US has really moved to what I might call the critique approach in this area. No one is willing to propose actual solutions, but everyone likes to complain and critique. This creates somewhat of a do nothing or can't do effect in govt especially and I think also reduces attractiveness of professions like policing or working in govt (you can't ever actually do or even try things without folks eagerly slamming you).
Pretty pathetic - and doesn't bode well for our covid testing / tracing response either.
The point is that the technology needs data. If crime is very low it is difficult to gather enough meaningful data.
But perhaps a counterpoint is that if crime is high maybe there is no need to use algorithms to tell you which areas are worse...
Overall, I'm thinking that the value of this technology is to use enough data to detect any trends or patterns that are not obvious and would likely go undetected otherwise.
But where the needle started at can absolutely matter. Ten degrees of movement need not be equal effort when it is obtained starting at 5 versus starting at 50.
The headroom that remains to improve will likely matter too, less the headroom more difficult it might be to reap the last inch.
The the software is named precobs in reference to minority reports precogs let's me believe their marketing team is hired from people selling penis enlargement on the internet. Perhaps they are just smarter because they know how often people are inclined to buy bullshit.