[Photon] location bias

[Holger Bruch]

Hi Peter,

indeed, submitting multiple requests (streets/addresses/transit stops
with distance_sort=true, cities with distance_sort=false) and presenting
them in a categorized drop down is the way I worked around this for now.
It is not optimal as on the client side, there is no information (like
score or importance) to merge the results in a meaningful way. And
results look strange if for one category exact matches are retrieved,
for the another only fuzzy one. Further, sorting by distance does not
take relevance into account. So providing a better scoring is preferable
in the first place, IMO.

Looking at the current location bias function, I think it does not
spread enough (boosts between 0.5 and 2) and it is not scale/viewport
dependent.

Elasticsearch's built-in decay-functions might penalize distant matches
too much, as they approach zero for distant matches:

https://www.elastic.co/guide/en/elasticsearch/guide/current/decay-functions.html

But an interesting feature they provide is the offset: For locations
closer than the offset the maximum location score is returned. If a
viewport or zoomlevel is provided as query parameter, this could be used
to return results closer to user expectations (at least mine): If
current map display is zoomed out, I'd like to preferably find matches
in the displayed bounding box , not only close to the center (but not
exclusivly, in the bounding box, so bounding box filtering is no option
here).

Regards,
Holger

Am 15.01.2018 um 09:00 schrieb Peter:
> Hi,
>
> When upgrading from the old to the new geocoding I can see a good
> average speed improvement overall.
>
> One bigger headache I have is that the location bias previously wasn't
> that well working and now is very strict. Here is an example
> illustrating the necessary different strengths. Always assume I use
> "berlin" as location bias (e.g. from a guessed location via IP address).
> Now if I use the location bias and
>
> 1. I type "railway station" (or "bahnhof") then I want stations of this
> city as first result, but
>
> 2. if I type another city like Paris I usually want a mix of "Berlin,
> Pariser Platz" and results from the city "Paris".
>
> Currently the location bias will only work in the first scenario. We
> could reduce the strength until it works but other use cases might
> require a different value, but I doubt that we'll improve the situation
> if we let the user tune this strength.
>
> Maybe I have to mix the results from two requests (one with the location
> bias, one without) on the client side? What do you think?
>
> Regards
> Peter
>