Latest revision as of 22:58, 27 April 2017
Metric Trees
Metric tree in an indexing structure that allows for efficient KNN search
Metric tree organizes a set of points hierarchically
- It's a binary tree: nodes = sets of points, root = all points
- sets across siblings (nodes on the same level) are all disjoint
- at each internal node all points are partitioned into 2 disjoint sets
Notation:
- let $N(v)$ be all points at node $v$
- $\text{left}(v), \text{right}(v)$ - left and right children of $v$
Splitting a node:
- choose two pivot points $p_l$ and $p_r$ from $N(v)$
- ideally these points should be selected s.t. the distance between them is largest:
- $(p_l, p_r) = \operatorname{arg max}\limits_{p_1, p_2 \in N(v)} \| p_1 - p_2 \|$
- but it takes $O(n^2)$ (where $n = |N(v)|$) to find optimal $p_l, p_r$
- heuristic:
- pick a random point $p \in N(v)$
- then let $p_l$ be point farthest from $p$
- and then let $p_r$ be point farthest from $p_l$
- once we have $(p_l, p_r)$ we can partition:
- project all points onto a line $u = p_r - p_l$
- find the median point $A$ along the line $u$
- all points on the left of $A$ got to $\text{left}(v)$, on the right of $A$ - to $\text{right}(v)$
- by using the median we ensure that the depth of the tree is $O(\log N)$ where $N$ is the total number of data points
- however finding the median is expensive
- heuristic:
- can use the mean point as well, i.e. $A = (p_l + p_r) / 2$
- let $L$ be a $d - 1$ dimensional plane orthogonal to $u$ that goes through $A$
- this $L$ is a decision boundary - we will use it for querying
After metric tree is constructed at each node we have:
- the decision boundary $L$
- a sphere $\mathbb B$ s.t. all points in $N(v)$ are in this sphere
- let $\text{center}(v)$ be the center of $\mathbb B$ and $r(v)$ be the radius
- so $N(v) \subseteq \mathbb B\big(\text{center}(v), r(v)\big)$
MT-DFS($q$) - the search algorithm
- search in a Metric Tree is a guided Depth-First Search
- the decision boundary $L$ at each node $n$ is used to decide whether to go left or right
- if $q$ is in the left, then go to $\text{left}(v)$, otherwise - to $\text{right}(v)$
- (or can project the query point to $u$, and then check if $q < A$ or not)
- all the time we maintain $x$: nearest neighbor found so far
- let $d = \| x - q \|$ - distance from best $x$ so far to the query
- we can use $d$ to prune nodes: we can check if a node is good or no point can better than $x$
- no point is better than $x$ if $\| \text{center}(r) - q \| - r(v) \geqslant d$
This algorithm is very efficient when dimensionality is $\leqslant 30$
- but slows down when it increases
Observation:
- MT often finds the NN very quickly and then spends 95% of the time verifying that this is the true NN
- can reduce this time with Spill-Trees
Spill-Trees
A Spill-Tree is a variant of Metric Tree
- in which children of a node can "spill over" onto each other
- i.e. $\text{left}(v)$ and $\text{right}(v)$ are no longer required to be disjoint
Partitioning
- the decision boundary $L$ still goes though $A$
- but we define two additional separate planes $L_L$ and $L_R$
- let $\tau$ be the area that both left and right children of $v$ can share
- $L_L = L - \tau$ and $L_R = L + \tau$
- $\tau$ is the size of overlap
- then $\text{left}(v)$ contains all points on the left of $L_R$ and $\text{right}(v)$ contains all the points on the right of $L_L$
- illustration:
Why allowing overlap?
- find the answer approximately, not exactly
SP-Search($q$)
- don't backtrack at all - just do a tree descent, not DFS
- consider a case when $q$ is close to $L$: it's true that the true NN might be on the other side of $L$
- so by allowing overlap we hope to catch the true NN on the over side
- and by varying $\tau$ we can reduce the probability of a mistake
Hybrid Spill-Trees
Problems with Spill-Trees: depth varies a lot with $\tau$
Construction
- let $\rho < 1$ be the balance threshold (usually $\rho = 0.7$)
- Similar to SP-Trees, but
- if either of $v$'s children contains more than $\rho \cdot | N(v) |$ elements
- then don't do overlapping nodes - use usual MT split and mark the node as "non-overlapping"
- this way we still can maintain the logarithmic depth
Search:
- also hybrid of both
- if non-overlapping: do backtracking
- if overlapping: don't backtrack
Both SP and MT aren't very efficient for $D \geqslant 30$
- but by Johnson-Lidenstrauss Lemma (see Achlioptas2003) know that
- we can always embed $N$ points into a subspace with dimensionality $\log N$
- with little distortion on pair-wise distances.
- so let's do a very simple embedding: Random Projections
- pick up a random subspace $S$ and project all data on $S$
So,
- do Random Projection as a preprocessing step: project all data points on $S$ and build the tree on the low dimensional representation
- by doing projection we'll lose some accuracy
- can fix that by doing multiple different random projections and do a hybrid search for each resulting tree separately
- if probability of failing to find the true NN is $\delta$, then do doing $L$ different projections we reduce this probability to $\delta^L$
References
- Uhlmann, Jeffrey K. "Metric trees." 1991. [1]
- Omohundro, Stephen M. "Bumptrees for efficient function, constraint, and classification learning." 1991. [2]
- Achlioptas, Dimitris. "Database-friendly random projections: Johnson-Lindenstrauss with binary coins." 2003. [3]
Sources
- Liu, Ting, et al. "An investigation of practical approximate nearest neighbor algorithms." 2004. [4]