Interface MatchEngine

  • All Known Implementing Classes:
    AbstractCartesianMatchEngine, AbstractSkyMatchEngine, AnisotropicCartesianMatchEngine, CombinedMatchEngine, CuboidCartesianMatchEngine, EllipseCartesianMatchEngine, EllipseSkyMatchEngine, EqualsMatchEngine, ErrorCartesianMatchEngine, ErrorSkyMatchEngine, FixedSkyMatchEngine, IsotropicCartesianMatchEngine, SphericalPolarMatchEngine

    public interface MatchEngine
    Defines the details of object matching criteria. This interface provides methods for ascertaining whether two table rows are to be linked - this usually means that they are to be assumed to refer to the same object. The methods act on 'tuples' - an array of objects defining the relevant characteristics of a row. Of course these tuples have to be prepared with understanding of what a particular implementation of this interface knows how to deal with, which can be obtained from the getTupleInfos() method. Typically a tuple will be a list of coordinates, such as RA and Dec.

    The business end of the interface consists of two methods. One tests whether two tuples count as matching or not, and assigns a closeness score if they are (in practice, this is likely to compare corresponding elements of the two submitted tuples allowing for some error in each one). The second is a bit more subtle: it must identify a set of bins into which possible matches for the tuple might fall. For the case of coordinate matching with errors, you would need to chop the whole possible space into a discrete set of zones, each with a given key, and return the key for each zone near enough to the submitted tuple (point) that it might contain a match for it.

    Formally, the requirements for correct implementations of this interface are as follows:

    1. matchScore(t1,t2) == matchScore(t2,t1)
    2. matchScore(t1,t2)>=0 implies a non-zero intersection of getBins(t1) and getBins(t2)
    The best efficiency will be achieved when:
    1. the intersection of getBins(t1) and getBins(t2) is as small as possible for non-matching t1 and t2 (preferably 0)
    2. the number of bins returned by getBins is as small as possible (preferably 1)
    These two efficiency requirements are usually conflicting to some extent.

    It may help to think of all this as a sort of fuzzy hash.

    Author:
    Mark Taylor (Starlink)
    • Field Detail

      • NO_BINS

        static final java.lang.Object[] NO_BINS
        Convenience constant - it's a zero-length array of objects, suitable for returning from getBins(java.lang.Object[]) if no match can result.
    • Method Detail

      • getBins

        java.lang.Object[] getBins​(java.lang.Object[] tuple)
        Returns a set of keys for bins into which possible matches for a given tuple might fall. The returned objects can be anything, but should have their equals and hashCode methods implemented properly for comparison.
        Parameters:
        tuple - tuple
        Returns:
        set of bin keys which might be returned by invoking this method on other tuples which count as matches for the submitted tuple
      • matchScore

        double matchScore​(java.lang.Object[] tuple1,
                          java.lang.Object[] tuple2)
        Indicates whether two tuples count as matching each other, and if so how closely. If tuple1 and tuple2 are considered as a matching pair, then a non-negative value should be returned indicating how close the match is - the higher the number the worse the match, and a return value of zero indicates a 'perfect' match. If the two tuples do not consitute a matching pair, then a negative number (conventionally -1.0) should be returned. This return value can be thought of as (and will often correspond physically with) the distance in some real or notional space between the points represented by the two submitted tuples.

        If there's no reason to do otherwise, the range 0..1 is recommended for successul matches. However, if the result has some sort of physical meaning (such as a distance in real space) that may be used instead.

        Parameters:
        tuple1 - one tuple
        tuple2 - the other tuple
        Returns:
        'distance' between tuple1 and tuple2; 0 is a perfect match, larger values indicate worse matches, negative values indicate no match
      • getMatchScoreInfo

        ValueInfo getMatchScoreInfo()
        Returns a description of the value returned by the matchScore(java.lang.Object[], java.lang.Object[]) method. The content class should be numeric (though need not be Double), and the name, description and units should be descriptive of whatever the physical significance of the value is. If the result of matchScore is not interesting (for instance, if it's always either 0 or -1), null may be returned.
        Returns:
        metadata for the match score results
      • getScoreScale

        double getScoreScale()
        Returns a scale value for the match score. The intention is that the result of matchScore/getScoreScale() is of order unity, and is thus comparable between different match engines.

        As a general rule, the result should be the maximum value ever returned from the matchScore method, corresponding to the least good successful match. For binary MatchEngine implementations (all matches are either score=0 or failures) a value of 1 is recommended. If nothing reliable can be said about the scale, NaN may be returned.

        Returns:
        scale of successful match scores, a positive finite number or NaN
      • getTupleInfos

        ValueInfo[] getTupleInfos()
        Returns a set of ValueInfo objects indicating what is required for the elements of each tuple. The length of this array is the number of elements in the tuple. Each element should at least have a defined name and content class. The info's nullable attribute has a special meaning: if true it means that it makes sense for this element of the tuple to be always blank (for instance assigned to no column).
        Returns:
        array of objects describing the requirements on each element of the tuples used for matching
      • getMatchParameters

        DescribedValue[] getMatchParameters()
        Returns a set of DescribedValue objects whose values can be modified to modify the matching criteria. Typically at least one of these will be some sort of tolerance separation which determines how close tuples must be to count as a match. This match engine's behaviour can be modified by calling DescribedValue.setValue(java.lang.Object) on the returned objects.
        Returns:
        array of described values which influence the match
      • getTuningParameters

        DescribedValue[] getTuningParameters()
        Returns a set of DescribedValue objects whose values can be modified to tune the performance of the match. This match engine's performance can be influenced by calling DescribedValue.setValue(java.lang.Object) on the returned objects.

        Changing these values will make no difference to the output of matchScore(java.lang.Object[], java.lang.Object[]), but may change the output of getBins(java.lang.Object[]). This may change the CPU and memory requirements of the match, but will not change the result. The default value should be something sensible, so that setting the value of these parameters is not in general required.

        Returns:
        array of described values which may influence match performance
      • getMatchBounds

        NdRange getMatchBounds​(NdRange[] inRanges,
                               int index)
        Given a range of tuple values, returns a range outside which no match to anything within that range can result. If the tuples on which this engine works represent some kind of space, the input values and output values specify a hyper-rectangular region of this space. In the common case in which the match criteria are based on proximity in this space up to a certain error, this method should return a rectangle which is like the input one but broadened in each direction by an amount corresponding to the error.

        Both the input and output rectangles are specified by tuples representing its opposite corners; equivalently, they are the minimum and maximum values of each tuple element. In either the input or output min/max tuples, any element may be null to indicate that no information is available on the bounds of that tuple element (coordinate).

        An array of n-dimensional ranges is given, though only one of them (specified by the index value) forms the basis for the output range. The other ranges in the input array may in some cases be needed as context in order to do the calculation. If the match error is fixed, only the single input n-d range is needed to work out the single output range. However, if the errors are obtained by looking at the tuples themselves (match errors are per-row) then in general the broadening has to be done using the maximum error of any of the tables involved in the match, not just the one to be broadened. For a long time, I didn't realise this, so versions of this software up to STIL v3.0-14 (Oct 2015) were not correctly broadening these ranges, leading to potentially missed associations near the edge of bounded regions.

        This method can be used by match algorithms which know in advance the range of coordinates they will match against and wish to reduce workload by not attempting matches which are bound to fail.

        For example, a 1-d Cartesian match engine with an isotropic match error 0.5 would turn input values of ((0,200),(10,210)) into output values ((-0.5,199.5),(10.5,210.5)).

        This method will only be called if canBoundMatch() returns true. Thus engines that cannot provide any useful information along these lines (for instance because none of its tuple elements is Comparable) do not need to implement it in a meaningful way.

        Parameters:
        inRanges - array of input ranges for the tables on which the match will take place; each element bounds the values for each tuple element in its corresponding table in a possible match (to put it another way - each element gives the coordinates of the opposite corners of a tuple-space rectangle covered by one input table)
        index - which element of the inRanges array for which the broadened output value is required
        Returns:
        output range, effectively inRanges[index] broadened by errors
        See Also:
        canBoundMatch()