After dinner in downtown San Francisco, I mentioned goodbye to buddies and pulled out my telephone to determine how one can get house. It was near 11:30 pm, and Uber estimates had been unusually lengthy. I opened Google Maps and checked out strolling instructions as a substitute. The routes had been comparable in distance, however I hesitated — not due to how lengthy the stroll would take, however as a result of I wasn’t certain how totally different elements of the route would really feel at the moment of night time. Google Maps might inform me the quickest manner house, nevertheless it couldn’t assist reply the query I used to be really asking: how can I filter for a route that takes me by means of safer blocks slightly than the quickest route?
Defining the Downside Assertion
Given a beginning location, ending location, day of the week, and time how can we predict the anticipated threat on the given strolling route? For instance, if I wish to stroll from the Ferry Constructing to Decrease Nob Hill, Google Maps reveals me the next route(s):
At a excessive stage, the issue I needed to unravel was this: given a beginning location, ending location, time of day, and day of week, how can we estimate the anticipated threat alongside a strolling route?
For instance, if I wish to stroll from Chinatown to Market & Van Ness, Google Maps presents a pair route choices, all taking roughly 40 minutes. Whereas it’s helpful to match distance and period, it doesn’t assist reply a extra contextual query: which elements of those routes are inclined to look totally different relying on the time I’m making the stroll? How does the identical route examine at 9 am on a Tuesday versus 11 pm on a Saturday?
As walks get longer — or cross by means of areas with very totally different historic exercise patterns — these questions develop into tougher to reply intuitively. Whereas San Francisco will not be uniquely unsafe in comparison with different main cities, public security continues to be a significant consideration, particularly when strolling by means of unfamiliar areas or at unfamiliar occasions. My objective was to construct a device for locals and guests alike that provides context to those selections — utilizing historic knowledge and machine studying to floor how threat varies throughout area and time, with out lowering the town to simplistic labels.
Getting the Knowledge + Pre-Processing
Fetching the Uncooked Dataset
The San Francisco Metropolis and County Departments publish police incident studies each day by means of the San Francisco Open Knowledge portal. The dataset spans from January 1, 2018 to the current and consists of structured info reminiscent of incident class, subcategory, description, time, and placement (latitude and longitude).
Categorizing Incidents Reported
One rapid problem with this knowledge is that not all incidents signify the identical stage or sort of threat. Treating all studies equally would blur significant variations — for instance, a minor vandalism report shouldn’t be weighted the identical as a violent incident. To handle this, I first extracted all distinctive mixtures of incident class, subcategory, and outline, which resulted in just a little over 800 distinct incident triplets.
Relatively than scoring particular person incidents immediately, I used an LLM to assign severity scores to every distinctive incident sort. This allowed me to normalize semantic variations within the knowledge whereas conserving the scoring constant and interpretable. Every incident sort was scored on three separate dimensions, every on a 0–10 scale:
- Hurt rating: the potential threat to human security and passersby
- Property rating: the potential for injury or lack of property
- Public disruption rating: the extent to which an incident disrupts regular public exercise
These three scores had been later mixed to type an total severity sign for every incident, which might then be aggregated spatially and temporally. This method made it doable to mannequin threat in a manner that displays each the frequency and the nature of reported incidents, slightly than counting on uncooked counts alone.
Geospatial Illustration
Offering uncooked latitude and longitude numbers won’t add a lot worth to the ML mannequin as a result of I have to group mixture incident context at a block and neighborhood-level. I wanted a technique to map a block or neighborhood to a set index to simplify function engineering and construct a constant spatial mapping. Minimize to the seminal engineering weblog revealed by Uber — H3.
Uber’s H3 blog describes how projecting an icosahedron (20-faced polyhedron) to the floor of the earth and hierarchically breaking them down into hexagonal shapes (and 12 strategically positioned pentagons) might help tessellate your complete map. Hexagons are particular as a result of there are one of many few common polygons that type common tessellations and its centerpoint is equidistant to it’s neighbors’, which simplifies smoothing over gradients.
The web site https://clupasq.github.io/h3-viewer/ is a enjoyable experiment to see what your location’s H3 Index is!
Temporal Illustration
Time is simply as essential as location when modeling strolling threat. Nevertheless, naïvely encoding hour and day as integers introduces discontinuities — 23:59 and 00:00 are numerically far aside, despite the fact that they’re solely a minute aside in actuality.
To handle this, I encoded time of day and day of week utilizing sine and cosine transformations, which signify cyclical values on a unit circle. This enables the mannequin to study that late-night and early-morning hours are temporally adjoining, and that days of the week wrap naturally from Saturday again to Sunday.
As well as, I aggregated incidents into 3-hour time home windows. Shorter home windows had been too sparse to provide dependable indicators, whereas bigger home windows obscured significant variations (for instance, early night versus late night time). Three-hour buckets struck a stability between granularity and stability, leading to intuitive intervals reminiscent of early morning, afternoon, and late night.
Remaining Characteristic Illustration
After preprocessing, every knowledge level consisted of:
- An H3 index representing location
- Cyclically encoded hour and day options
- An aggregated severity sign derived from historic incidents
The mannequin was then educated to foretell the anticipated threat for a given H3 cell, at a given time of day and day of week. In apply, which means when a consumer opens the app and gives a location and time, the system has sufficient context to estimate how strolling threat varies throughout close by blocks.
Coaching the Mannequin Utilizing XGBoost
Why XGBoost?
With the geospatial and temporal options prepared, I knew I wanted to leverage a mannequin which might seize non-linear patterns within the dataset whereas offering low latency to carry out inference on a number of segments in a route. XGBoost was a pure match for a pair causes:
- Tree-based fashions are naturally sturdy at modeling heterogenous knowledge — categorical spatial indices, cyclical time options, and sparse inputs can coexist with out heavy function scaling or normalization
- Characteristic results are extra interpretable than in deep neural networks, which are inclined to introduce pointless opacity for tabular knowledge.
- Flexibility in targets and regularization made it doable to mannequin threat in a manner that aligns with the construction of the issue.
Whereas I did think about alternate options reminiscent of linear fashions, random forests, and neural networks, they had been unsatisfactory attributable to lack of ability to seize nuance in knowledge, excessive latency at inference time, or over-complication for tabular knowledge. XGBoost strikes the very best stability between efficiency and practicality.
Modeling Anticipated Danger
It’s essential to make clear earlier than we transfer on that modeling anticipated threat will not be a Gaussian drawback. When modeling incident charges within the metropolis, I seen per [H3, time] cell that:
- a number of cells have incident depend = 0 and/or whole threat = 0
- a number of cells have simply 1–2 incidents
- handful cells have incidents many incidents (> 1000)
- excessive occasions happen, however hardly ever
These are indicators that my mannequin is neither symmetrical nor will the information factors cluster round a set imply. These properties instantly rule out widespread assumptions like usually distributed errors.
That is the place Tweedie regression turns into helpful.
What’s Tweedie Regression?
Put merely, Tweedie regression says: “Your worth is the sum of random occasions the place the variety of occasions is random and every occasion has a optimistic random measurement.” This suits the crime incident mannequin completely.
Tweedie regression combines Poisson and Gamma distribution processes to mannequin the variety of incidents and the scale (threat rating) of every incident. For example:
- Poisson course of: in window 6pm-9pm on December tenth, 2025 what number of incidents occurred in H3 index 89283082873ffff?
- Gamma distribution: how extreme was every occasion that occurred in H3 index 89283082873ffff between 6pm-9pm on December tenth, 2025?
Why This Issues?
A concrete instance from the information illustrates why this framing is essential.
Within the Presidio, there was a single, uncommon high-severity incident that scored near 9/10. In distinction, a block close to 300 Hyde Road within the Tenderloin has 1000’s of incidents over time, however with a decrease common severity. Tweedie breaks it down as:
Anticipated threat = E[#incidents] × E[severity]# Presidio
E[#] ≈ ~0
E[severity] = excessive
→ Anticipated threat ≈ nonetheless ~0
# Tenderloin
E[#] = excessive
E[severity] = medium
→ Anticipated threat = giantSubsequently, if the high-risk occasions are inclined to occur extra usually in Presidio then it’ll modify the anticipated threat accordingly and lift the output scores. Tweedie handles the goal’s zero‑heavy, proper‑skewed distribution and the enter options we mentioned earlier simply clarify variation in that focus on.
Framing the Final result
The result’s a mannequin that predicts anticipated threat, not conditional severity and never binary security labels. This distinction issues. It avoids overreacting to uncommon however excessive occasions, whereas nonetheless reflecting sustained patterns that emerge over time.
Remaining Steps + Deployments
To deliver the mannequin to life, I used the Google Maps API to construct an internet site which integrates the maps, routes, and route UI on which I can overlay colours primarily based on the chance scores. I color-coded the segments by taking a percentile of distributions in my knowledge, i.e. rating ≤ P50 = inexperienced (protected), rating ≤ P75 = yellow (reasonably protected), rating ≤ P90 = orange (reasonably dangerous), else crimson (dangerous). I additionally added a logic to re-route the consumer by means of a safer route if the detour will not be over 15% of the unique period. This may be tweaked, however I left it as is for now since with the San Francisco hills a 15% detour might work you a large number.
I additionally deployed backend on Render and frontend on Vercel.
Placing StreetSense To Use!
And now, going again to the primary instance we checked out — the journey from Chinatown to Market & Van Ness, however now with our new mannequin + software we’ve constructed!
Right here’s how the stroll appears to be like like at 9am on a Tuesday versus 11pm on a Saturday:
Within the first picture, the segments in Chinatown that are inexperienced have a decrease incident and severity depend in comparison with segments that are crimson and the information backs it too. The cool half concerning the second picture is that it routinely re-routes the consumer by means of a route which is safer at 11pm on a Saturday night time. That is the form of contextual decision-making I initially wished for — and the motivation behind constructing StreetSense.
Remaining Ideas and Potential Enhancements
Whereas the present system captures spatial and temporal patterns in historic incidents, there are clear areas for enchancment:
- incorporating real-time indicators
- utilizing additional floor reality knowledge to validate and practice
(a) if an incident was marked as 4/10 threat rating for theft and we will discover out by means of the San Francisco database that an arrest was made, we will bump it as much as a 5/10 - make H3 index delicate to neighboring cells — Outer Richmond ~ Central Richmond so the mannequin ought to infer proximity and contextual info must be partially shared
- Develop spatial options past H3 ID (neighbor aggregation, distance to hotspots, land‑use options).
- deeper exploration of various strategies of dealing with of incident knowledge + evaluations
(a) experiment with totally different XGBoost goal features reminiscent of Pseudo Huber Loss
(b) Leverage hyperparameter optimization frameworks and consider totally different mixtures of values
(c) experiment with neural networks - increasing past a single metropolis would all make the mannequin extra sturdy
Like several mannequin constructed on historic knowledge, StreetSense displays previous patterns slightly than predicting particular person outcomes, and it must be used as a device for context slightly than certainty. Finally, the objective is to not label locations as protected or unsafe, however to assist individuals make extra knowledgeable, situationally conscious decisions as they transfer by means of a metropolis.
Strive StreetSense: https://san-francisco-safety-index.vercel.app/
Knowledge Sources & Licensing
This mission makes use of publicly out there knowledge from the San Francisco Open Knowledge Portal:
- San Francisco Police Division Incident Stories
– Supply: San Francisco Open Knowledge Portal (https://data.sfgov.org)
– License: Open Knowledge Commons Public Area Dedication and License (PDDL)
All datasets used are publicly out there and permitted for reuse, modification, and industrial use below their respective open knowledge licenses.
Acknowledgement & References
I’d wish to thank the San Francisco Open Knowledge group for sustaining high-quality public datasets that make initiatives like this doable.
Extra references that knowledgeable my understanding of the strategies and ideas used on this work embrace:
