R-Trees
Accessing Spatial Data
In the beginning
The B-Tree provided a foundation for RTrees. But whats a B-Tree?
A data structure for storing sorted data with
amortized run times for insertion and deletion
Often used for data stored on long latency I/O
(filesystems and DBs) because child nodes
can be accessed together (since they are in
order)
B-Tree
From wikipedia
Whats wrong with B-Trees
B-Trees cannot store new types of data
Specifically people wanted to store
geometrical data and multi-dimensional data
The R-Tree provided a way to do that (thanx
to Guttman 84)
R-Trees
R-Trees can organize any-dimensional data
by representing the data by a minimum
bounding box.
Each node bounds its children. A node can
have many objects in it
The leaves point to the actual objects (stored
on disk probably)
The height is always log n (it is height
balanced)
R-Tree Example
From http://lacot.org/public/enst/bda/img/schema1.gif
Operations
Searching: look at all nodes that intersect,
then recurse into those nodes. Many paths
may lead nowhere
Insertion: Locate place to insert node through
searching and insert.
If a node is full, then a split needs to be done
Deletion: node becomes underfull. Reinsert
other nodes to maintain balance.
Splitting Full Nodes
Linear choose far apart nodes as ends.
Randomly choose nodes and assign them so
that they require the smallest MBR
enlargement
Quadratic choose two nodes so the dead
space between them is maximized. Insert
nodes so area enlargement is minimized
Exponential search all possible groupings
Note: Only criteria is MBR area enlargement
Demo
How can we visualize the R-Tree
By clicking here
Variants - R+ Trees
Avoids multiple paths during searching.
Objects may be stored in multiple nodes
MBRs of nodes at same tree level do not overlap
On insertion/deletion the tree may change
downward or upward in order to maintain the
structure
R+ Tree
http://perso.enst.fr/~saglio/bdas/EPFL0525/sld041.htm
Variants: Hilbert R-Tree
Similar to other R-Trees except that the Hilbert value
of its rectangle centroid is calculated.
That key is used to guide the insertion
On an overflow, evenly divide between two nodes
Experiments has shown that this scheme
significantly improves performance and decreases
insertion complexity.
Hilbert R-tree achieves up to 28% saving in the
number of pages touched compared to R*-tree.
Hilbert Value??
The Hilbert value of an object is found by
interleaving the bits of its x and y coordinates, and
then chopping the binary string into 2-bit strings.
Then, for every 2-bit string, if the value is 0, we
replace every 1 in the original string with a 3, and
vice-versa.
If the value of the 2-bit string is 3, we replace all 2s
and 0s in a similar fashion.
After this is done, you put all the 2-bit strings back
together and compute the decimal value of the
binary string;
This is the Hilbert value of the object.
http://www-users.cs.umn.edu/research/shashi-group/CS8715/exercise_ans.doc
R*-Tree
The original R-Tree only uses minimized
MBR area to determine node splitting.
There are other factors to consider as well
that can have a great impact depending on
the data
By considering the other factors, R*-Trees
become faster for spatial and point access
queries.
Problems in original R-Tree
Because the only criteria is to minimize area
1.
2.
Certain types of data may create small areas but
large distances which will initiate a bad split.
If one group reaches a maximum number of
entries, the rest of assigned without
consideration of their geometry.
Greene tried to solve, but he only used the
split axis more criteria needs to be used
Splitting overfilled nodes
Why is this overfull?
R*-Tree Parameters
1.
2.
3.
4.
Area covered by a rectangle should be
minimized
Overlap should be minimized
The sum of the lengths of the edges
(margins) should be minimized
Storage utilization should be maximized
(resulting in smaller tree height)
Splitting in R*-Trees
1)
2)
3)
4)
Entries are sorted by their lower value, then
their upper value of their rectangles. All
possible distributions are determined
Compute the sum of the margin values and
choose the axis with the minimum as the
split axis
Along the split axis, choose the distribution
with the minimum overlap
Distribute entries into these two groups
Deleting and Forced Re-insertion
Experimentally, it was shown that re-inserting
data provided large (20-50%) improvement in
performance.
Thus, randomly deleting half the data and reinserting is a good way to keep the structure
balanced.
Results
Lots of data sets and lots of query types.
One example: Real Data: MBRs of elevation lines.
100K objects
Disk accesses
Query
Storage util. On insert
After build up
RC-Trees
Changing motivations:
Memory large enough to store objects
Its possible to store the object geometry and not
just the MBR representation.
Data is dynamic and transient
Spatial objects naturally overlap (ie: stock market
triggers)
RC-Trees
Take advantage of dynamic segmentation
If the original geometry is thrown away, then
later on the MBR cannot be modified to
represent new changes to the tree
RC Tree does
1.
2.
3.
Clipping
Domain Reduction
Rebalancing
Discriminators
A discriminator is used to decide (in binary) which
direction a node should go in. (It means its a binary
tree, unlike other R-Trees)
It partitions the space
If an object intersects a discriminator, the object can
be clipped into two parts
When an object is clipped, the space it takes up (in
terms of its MBR) is reduced (aka domain reduction)
This allows for removal of dead space and faster
point query lookups
Domain Reduction and Clipping
Operations
Insert, Delete and Search are straightforward
What happens on an node that has been
overflowed?
Choose a discriminator to partition the object
into balanced sets
How is a discriminator chosen?
Partitioning
Two methods for finding a discriminator for a
partition
RC-MID faster, but ignores balancing and
clipping. Uses pre-computed data to
determine and average discriminator.
Problems?
Different distributions greatly affect partition
Space requirements can be huge
Partitioning Take 2
RC-SWEEP
sorts objects.
Candidates for discriminators are the boundaries
of the MBRs
Assign a weight to each candidate using a
formula not shown here
Choose the minimum
Problems?
Slower, but space costs much better than RC-MID
(which keeps info about nodes)
Rebuilding
The tree can take a certain degree of
flexibility in its structure before needing to be
rebalanced
On an insert, check if the height is too
imbalanced
If so, go to the imbalanced subtree and flush
the items, sort and call split on them to get a
better balancing
Experimentation
CPU execution time not a good measure. (although
they still calculate it)
Instead use number of discriminators compared
Lots of results
Result summary:
Insertion a little more expensive (because of possible
rebalancing)
Querying for point or spatial data faster (and fewer memory
accesses) than all previous incarnations
Storage requirements not that bad
Dynamic segmentation (ie recalculating MBRs) can help a
lot
Controlling space with factor (by disallowing further
splitting) controls space costs
