CHANGES

3.6.11

- BidirectionalImmutableGraph has been ported from the big version.

- BitStreamArcLabelledImmutableGraph has now copiable iterators.

3.6.10

- Now ScatteredArcsASCIIGraph propagates BatchGraph's copiable iterators.
  As a consequence, compression will happen in parallel when possible.

- Official support starts now at Java 9.

3.6.9

- Updated licensing info in the POM.

3.6.8

- Removed (almost) unused dependencies.

- Upgraded to newest versions of Apache Commons.

3.6.7

- OSGi metadata and default modularization.

- Manual reduction in strength in arithmetic operations everywhere.

3.6.6

- WebGraph is now dually licensed under the Lesser GNU Public License 2.1+
  or the Apache Software License 2.0.

- Added automatic module name.

3.6.5

- Efficiency improvements.

- Transform provides direct support to compute efficiently the simple version
  of a graph, either directly or via a precomputed transpose.

3.6.4

- Fixed SLF4J dependencies (again).

3.6.3

- Fixed build.xml: moved to Java 9 (following Sux4J), and now use the
  classifier in naming artifacts.

3.6.2

- New tool to export in GraphViz's .dot format.

- ScatteredArcsASCIIGraph can now accept zipped files. Thanks to Thibault
  Allançon for implementing this feature.

3.6.1

- Removed spurious debug print.

3.6.0

- Java 8-only.

- Fixed obscure bug in ShiftByOneArcListASCIIGraph: if the arc list
  was specified on the command line (no -1 option) and more than
  one core was available, the graph would have not been shifted.
  Thanks to Luca Prigioniero for reporting this bug.

3.5.3

- Fixed lack of implementation of loadMapped() in ImmutableSubgraph.
  Thanks to Massimo Santini and Pierlauro Sciarelli for reporting this bug.

3.5.2

- Removed last dependency from COLT. Unfortunately, ErdosRenyiGraph now
  uses a different binomial distribution generator, so generated graphs
  will be different, even using the same seed.

- New implementations by Michele Borassi of the SumSweep algorithm for
  diameter, radius and eccentricities of directed and undirected graphs.

3.5.1

- New TopKGeometricCentrality class by Michele Borassi.

3.5.0

- New mechanism for parallel compression based on the notion of "copiable
  iterators". Implemented in BVGraph and all derivative classes (e.g.,
  transposed graphs).

- HyperBall now accepts weights on the nodes.

3.4.3

- The family of loadSequential() methods have been deprecated, and
  replaced in code by loadOffline() or loadMapped().

3.4.2

- Fixed dependencies.

3.4.1

- Significantly improved performance of HyperBall on graphs with a highly
  skewed (e.g., heavy-tailed) outdegree distribution (e.g., transposed web
  graphs).

- Fixed wrong estimation of memory used.

- Now ConnectedComponents writes results using "wcc" instead of "scc".

3.4.0

- Fixed problem with obsolete BitSet used to store buckets in Stats.

- New parallel classes to compute geometric centralities and betweenness.

3.3.3

- Regressed to fastutil's quicksort calls in case of array fragments. Java
  7's Arrays.sort() has a memory bug that was killing the performance of a
  number of methods.

3.3.2

- We now distribute SpeedTest, hoping to improve the quality of benchmarks
  in the literature.

3.3.1

- Adapted to new DSI utilities.

3.3.0

- HyperBall sports a new adaptive decomposition scheme that 
  is based on the number of arcs to be scanned, rather than
  on the number of nodes.

- Fixed bug in the computation of the buckets. If you have used the new
  iterative implementation of Tarjan's algorithm
  (StronglyConnectedComponents) to compute buckets please recompute them.

3.2.1

- New iterative implementation of Tarjan's algorithm.

- HyperBall can now compute Nieminen's centrality.

3.2.0

- New selectable upper bound for EFGraph makes it possible to build
  "fake" graphs in whcih successors are greater than or equal to
  the number of nodes (this was already possible with BVGraph). Useful
  for incremental graph construction.

- New IncrementalImmutableSequentialGraph adapter, which provides an
  inversion of control for storing graphs: you supply, one at a time,
  the successor list of each node.

3.1.0

- New HyperBall implementation of the HyperANF idea. It computes several
  kind of geometric centrality and once in systolic local mode uses time
  proportional to the number of edges causing a modification, setting in
  practice the expected run time to the theoretical bound O(m log n).

- Now terminal nodes have closeness equal to zero (terminal nodes already
  had Lin's centrality equal to one).

- The DecimalFormat object used to print data is has now a fixed US locale.

- New EFGraph implementation (backported from the big version) using the
  Elias-Fano representation of monotone sequences. Compression is not so
  good, but successor enumeration is blazingly fast and the implementation
  returns a skippable iterator which provides constant-time search of
  nodes by lower bound.

- Both BVGraph and EFGraph have outdegree caching and exact unwrapping
  of successorArray(). This should bring performance improvements.

3.0.9

- We switched to SLF4J for logging.

3.0.8

- Now Webgraph includes an adapter towards the Jung graph analysis
  framework. The main method of the class can write immutable graphs
  into Pajek format.

3.0.7

- Now ScatteredArcsASCIIGraph accepts a translation function from
  node identifiers to node numbers.

3.0.5

- New ImmutableGraph.outdegrees() method that exposes the degrees of a graph
  as an IntIterator.

- RandomGraph removed.  

- ErdosRenyiGraph and almost all transformed graphs now support copy().

- New Transform.NodeClassFilter.

3.0.2

- A new class performs parallel breadth-first visits. NeighbourhoodFunction
  now uses it.

- A new class DoubleSweepFringeDiameter computes heuristically the
  diameter of symmetric graphs using parallel breadth-first visits.

- A new class ConnectedComponents computes the connected components of
  symmetric graphs using parallel breadth-virst visits.

- Fixed in ScatteredArcsASCIIGraph a bug (inherited from fastutil) that
  was generating spurious nodes for graphs with more than 100 million
  nodes.

- Moved unit tests to Junit4.

- Fixed bug in mapOffline() that was causing some spurious zero-degree
  nodes to be part of the graph if a tail of nodes of the original graph
  was erased.

- Fixed rare race condition in HyperANF external executions using less
  than 64 registers.

- New method to compute the median of all distances.

- HyperApproximateNeighbourhoodFunction now handles graphs that do not
  implement numArcs().

- Major revamp of ImmutableSubgraph, which now uses an additional integer
  per supergraph node, but it's much faster.

- New subclass of ImmutabeSubgraph that automatically builds a subgraph
  formed by nodes with outdegree in a specified range.

- ErdosRenyiGraph is now an ImmutableGraph.

- New class for checks.

- HyperApproximateNeighbourhoodFunction can now be reused by calling
  init(seed).

3.0.1

- We now try to adapt classes from the big version if possible.

- Offset deltas can now be long (in case you have really crazy graphs).

3.0

- WARNING: This release has minor binary incompatibilities with previous
  releases, mainly due to the move from the interface
  it.unimi.dsi.util.LongBigList to the now standard
  it.unimi.dsi.fasutil.longs.LongBigList. It is part of a parallel release
  of fastutil, the DSI Utilities, Sux4J, MG4J, WebGraph, etc. that were
  all modified to fit the new interface, and that prepare the way for our
  "big" versions, that is, supporting >2^31 entries in arrays (simulated),
  elements in lists, terms, documents, nodes, etc. Please read our (short)
  "Moving Java to Big Data" document (JavaBig.pdf) for details.

- We now require Java 6.

- New mapOffline method for mapping large graphs.

- All offline transformation methods now compress their batches; the
  resulting batch size is comparable to the size of the BVGraph
  representation with copying disabled.

- Now we never resort to the ImmutableGraph implementation of
  NodeIterator when iterating over an ImmutableSubgraph if
  the supergraph does not implement random access. We used to
  do it when the graph was very sparse, but not checking for
  random access was not a good idea.

- The computation of the ratio w.r.t. the information-theoretical
  lower bound (associated to the key "compratio" in the property file of a
  graph) was wrong and has been fixed.

- A number of classes deal with exact and approximate computation of
  the neighbourhood function of a graph, its distance density function,
  and derived values. See our paper about HyperANF (this stuff was actually
  introduced in 2.4.5, but we forgot to mention).

- Fixed an occasional infinite loop in ErdosRenyiGraph.

- New class for reading scattered arc lists (ids need to be contiguous).

- BVGraph.store() now sets up the node iterator before starting the
  progress logger. This should provide more sensible estimates of time to
  completion in case of offline methods.

- BVGraph node iterators have now a finalize() method that will close the
  underlying bit stream (and thus possibly the underlying file handle)
  when the iterator is no longer used.

- HyperApproximateNeighbourhoodFunction would not work with an offline
  graph even with a single thread (thanks to Lars Backstrom for reporting
  this bug).

- HyperApproximateNeighbourhoodFunction supports now 16 (-l4) or 32 (-l5)
  registers per counter.

- Fixed old-standing synchronization bug in
  HyperApproximateNeighbourhoodFunction.

- New static method NeighbourhoodFunction.harmonicDiameter().

2.4.5

- WebGraph is now distributed under the GNU General Public License 3.

- WARNING: A small modification to the coding makes it possible to
  compress graphs with more than 1B nodes (always up to 2B nodes). This,
  however, means that such graphs will not be readable by previous
  versions, which will crash. We felt this was not such a big issue, as
  such graphs were not previously compressible at all, so the version
  number has not been bumped.

- StronglyConnectedComponents no longer uses a separate thread to
  set the stack size. The process was not guaranteed (by contract)
  to set the stack size at all. The computation now runs in the main
  thread, and we suggest using suitable JVM options to set a large
  stack size.

- BVGraph now computes a wealth of statistical data related to the
  behaviour of the compression algorithm.

- A number of classes deal with estimating efficiently the neighbourhood
  function of a graph, the effective diameter, and the spid
  (shortest-paths index of dispersion).

- A caching mechanism has been put in place to make offsets loading
  orderds of magnitudes faster. You can generate a cached, serialised
  EliasFanoMonotoneLongBigList with the option -L of BVGraph, and then it
  will be loaded instead of scanning the offsets file.

- Fixed bug in the definition of in/out trees in ArrayListMutableGraph.

- Now Stats computes loops.

- BitStreamArcLabelledGraph was not supporting offset steps any longer,
  but constructors and static methods still made it possible to pass
  an offset step. This has been fixed.

- Some residual documentation about offset steps has been removed.

- A new cutoff option makes it possible to eliminate from a graph
  generated by a map operation on the command line (see Transform) all
  nodes whose index is too large. This is useful in conjunction with maps
  that quotient a graph (e.g., to get just large strongly connected
  components).

2.4.4

- The empty Formatter constructor was causing problems on localised systems.
  Now we use Locale.ROOT.

- offsetStep > 1 no longer exist. It is not necessary with the new Elias-Fano
  offset list.

- Speed improvements in random access to a BVGraph.

- Fixed semantics of ImmutableGraph.successorArray(): implementations are
  now forced to return a new array at each call. All implementations in
  WebGraph are now compliant.

- Now nodeIterator(int) in ImmutableSequentialGraph is implemented so that
  it calls nodeIterator() and then skips to the desired node.

- Fixed bad bug in UnionImmutableGraph: the node for which the cache was
  active was not set by successors().

- We now output some basic, exponentially binned stats for the distribution
  of successor gaps and residual gaps. From these data we also compute an
  approximation of the average gap for successors and for residuals.

- We now record how much space is used by every component of the compression
  algorithm.

- Following some research, the default minimum interval length in BVGraph
  is now 4.

2.4.3

- Fixed ArrayListMutableGraph.addNodes() (thanks to Erik Lumer for
  finding and fixing this bug).

- New options to shift the output of ASCII graphs.

- RemappedImmutableGraph.successorArray(x) was providing the same array on every
  call, thus making the inherited successors(x) method unusable to scan in
  parallel different lists. Fixed (now it returns a copy of the array, instead).

- New random transformation that permutes randomly a graph.

2.4.2

- Transform was not derelativising underlying-graph filenames.

- New classes to support flexible filtering of arc-labelled
  graphs. See the new action "larcfilter" of Transform and
  the interface LabelledArcFilter.

- StronglyConnectedComponents now uses a filter for labelled arcs, in case
  you want to compute components of a subgraph.

- Fixed old bug in StronglyConnectedComponents: the renumber
  option was not working.

- New Transform.compose() transformation that composes graphs
  (i.e., it computes the graph represented by the product
  of the Boolean matrices representing two graphs). You can
  even compose labelled graphs by providing a semiring to
  compose labels.

- Now label files can be longer than 2GiB.

2.4.1

- Fixed stupid null-pointer bug in BitStreamImmutableArcLabelledGraph.

2.4

- WARNING: There are more general relabelling strategies, but older
  code must be slightly refactored.

- Now BitStreamArcLabelledImmutableGraph supports contextual labels.
  They accept an additional directory as context, to resolve relative
  names.

2.3

- Fixed bug in BitStreamArcLabelledImmutableGraph: labels longer
  than 2Gi would have caused overflows.

- The new pointer loading system has been extended to arc-labelled graphs,
  too.

2.2

- New pointer loading system based on succinct representations. Now on
  typical web graphs pointers occupy 8-9 [sic] bits per element, thus
  almost halving the memory footprint.. The performance drop is about
  10-15% (measured in ns/link on an Opteron) for reference chains of length
  3 (and it decreases for shorter chains).

- New greyPerm transform to just get the permutation.

- ArcLabelledImmutableGraph now strengthens the implementation of
  nodeIterator() based on the random-access methods.

- Fixed lack of checks in integer key labels.

- New defensive check in BVGraph against badly implemented ImmutableGraphs.

2.1

- WARNING: Refactored to be based on dsiutils and Sux4J. This will cause
  some incompatibilities, in particular with loggers.

- Moved DocumentSequenceImmutableGraph to LAW, to avoid dependency
  on MG4J and vice versa.

2.0

- WARNING: WebGraph 2+ is not fully compatible with previous versions, and
  requires some minor refactoring: due to the new lazy architecture, the
  semantics of successors() has radically changed; in particular, a
  LazyIntIterator is returned instead of an IntIterator. Please refer to
  the ImmutableGraph documentation.

- New customised class parser that will prepend it.unimi.dsi.webgraph.
  and it.unimi.dsi.webgraph.labelling. to classes specified on 
  the command line (at last!).

- New ArcListASCIIGraph that specifies one arc per line and guesses
  the number of nodes. A special implementation can be used when
  nodes are numbered from one.

- New --spec switch that makes it possible to specify graphs as
  class names with arguments. Most useful to turn MG4J's document
  sequences into graphs using a VirtualDocumentResolver.

- Slightly relaxed contract for numNodes() (to make ArcListASCIIGraph
  conforming).

- New classes for union and transposition of labelled graphs. Transform
  has been adapted to use automatically BitStreamArcLabelledImmutableGraph
  to save arc-labelled graph, but the class is settable.

- Arc-labelled graphs must expose a prototype of their labels.

- New store() suggested methods for arc-labelled graphs.

- New Stats class for computing basic statistical data.

- Very, very, old bug in BVGraph has been fixed. nodeIterator(from)
  with from>1 was not working properly. Thanks to Francesco Zumpano
  and Pierluigi Origlia for finding this bug.

- New example class to interface your data with arc-labelled graph classes.

- Integer labels have a public value fields.

- Load methods of BVGraph now look for an offsetstep property to set
  the offset step externally.

- New extension for label offsets (.labeloffsets) and new property
  key for the underlying graph (underlyinggraph). Watch out!

- New relabelling wrapper to change the labels of a graph.

- New class implementing a variant of the Tarjan algorithm.

- All standard extensions and property keys are now defined by string constants.

- New algo package. We start with strongly connected components.

1.7

- Brand new ArcLabelledGraph 

- Deprecated classes and methods have been removed.

- Revamped OutdegreeStats class.

- New loadOnce() method for loading graphs on-the-fly. Very useful for
  generating an ASCIIGraph to standard output can compressing it without
  actually storing it.

- New randomAccess() method that tells you whether a 
  graph supports random access.

- A number of new packages containing unit tests.

- Fixed bug in ImmutableSubgraph: the property subgraphnodes
  was not actually read.

- Implemented a workaround for bug #6478546 (you can't do read() on large
  arrays when you have a lot of heap--bizarre, isn't it?).

1.6

- Most load() static methods now override the return type and
  declare the actual returned type, usually more specific (e.g.,
  BVGraph.load() returns a BVGraph).

- Graphs can now be transposed with an offline method. It is
  slower than the in-memory method, but it can transpose arbitrarily
  large graphs.

1.5

- IMPORTANT: WebGraph requires now Java 5.

- New ArrayListMutableGraph class that makes it easy to create
  dynamically graphs, and then exposes them as an ImmutableGraph.

- New documentation and example on how to import your data in
  WebGraph.

- All code moved from ProgressMeter to ProgressLogger. All old
  methods are deprecated.

- Command line parsing entirely handled by JSAP.

- The default maximum reference count for BVGraph is now 3.

- ASCIIGraph has been revamped to be usable to convert offline
  large graphs.

- The basename property was never used, and it is no longer saved.

1.4.1

- New method writeOffsets() and corresponding -O option in BVGraph
  which writes the offsets of a graph computing them from the graph
  representation (.graph file). This allows to distribute directly
  just the .graph and the .properties files.

- Incompatible ImmutableSubgraph, with more (hopefully) sensible 
  method names.

1.4

- Now various classes use the ImmutableGraph reflection methods.

- New ImmutableSubgraph class for storing and manipulating subgraphs
  holding just a reference to the node subset.

- New Transform static container with common constructions, and
  computation of Gray code ordering.

- Fixed lack of error message when accessing randomly successor
  in a sequentially loaded BVGraph.

1.2.4

- The graph class name is now obtained using getName(), and
  kluges have been placed that make also old graphs work.

- New explicit convention for storing the graph class name in a property file.

- New static methods in ImmutableGraph that load a graph using reflection
  and the convention above.

- Fixed lack of check or null pm.

- Fixed lack of loadOffline() method in BVGraph (causing infinite recursion).

1.2.2

- Aligned usage of iterators with fastutil 3.1.

1.2.1

- Fixed a stupid bug (in one case we forgot to reallocate a new
  FastMultiByteArrayInputStream).

- Fixed another stupid bug (using a standard, memory-stored
  graph would have not worked!).

1.2

- BVGraph now supports graphs larger than 2 GiB (in fact, up to 256 PiB)
  using (transparently) FastMultiByteArrayInputStream.

1.1

- The return type of the load method has been changed to ImmutableGraph,
  so to make it possible to override it in subclasses. This might require some
  additional type casting in existing code.

1.0r2

- Updated to new fastutil class set.

1.0

- First public release.