Re: [igraph] edge.betweenness returns negative or NaN values

[Szabolcs Horvát]

On 14 September 2015 at 00:30, Szabolcs Horvát <address@hidden> wrote:
> Just to have a simple test case:
>
> Based on a comparison with Mathematica, for an n*n grid graph things
> start to go wrong at n=35. For n=34 the edge betweenness is still
> accurate.
>

While the n=34 value is about right for the limits of betweenness
calculation without bigints, it happens to be the case that for n > 34
Mathematica also gives bad results (different ones).

> On 13 September 2015 at 22:37, Tamas Nepusz <address@hidden> wrote:
>> Ah, my bad, I did not notice that you meant the edge betweenness and
>> not the node betweenness. It looks like the edge betweenness related
>> code does not include a similar bigint workaround. This should be
>> fixed in the next version then - can you add a bug report to
>> https://github.com/igraph/igraph/issues ?
>>
>> Thanks,
>> T.
>>
>> T.
>>
>>
>> On Sun, Sep 13, 2015 at 9:24 AM, Vincent Labatut
>> <address@hidden> wrote:
>>> Hello Tamas,
>>>
>>> and thanks for your answer.
>>> I was thinking of something like that, since the networks are lattices and
>>> this case is mentionned in the documentation.
>>> I also had noticed the "nobigint" parameter for the "betweenness" functions,
>>> however it does not exist for the "edge.betweenness" functions.
>>> Best,
>>> Vincent
>>>
>>>> Hello,
>>>>
>>>> The negative values are most likely due to integer overflows (i.e. the
>>>> betweenness score would be too large and the underlying variable in
>>>> which igraph computes the betweenness score overflows). You can get
>>>> around this by passing nobigint=FALSE to the edge betweenness call -
>>>> it will make igraph use "big integers", which can hold arbitrarily
>>>> large numbers at the expense of being somewhat slower.
>>>>
>>>> As for the NaNs, it could be a bug, but let's see first whether the
>>>> issue persists with nobigint=FALSE. If so, let us know and try to post
>>>> a small example on which we could reproduce the issue with NaNs.
>>>>
>>>> T.
>>>>
>>>>
>>>>
>>>> On Sat, Sep 12, 2015 at 3:08 PM, Vincent Labatut
>>>> <address@hidden> wrote:
>>>>> Hello,
>>>>>
>>>>> I am processing the edge-betweenness of various networks using R igraph
>>>>> version 7.1. Those are spatial networks (each node has a (x,y) position)
>>>>> and
>>>>> I am using the "weight" option of the "edge.betweenness" function to take
>>>>> the spatial distances into account. This spatial distance is stored in an
>>>>> edge attribute called "dist".
>>>>>
>>>>> Here is the command I use:
>>>>> edge.betweenness(graph=g, weights=E(g)$dist)
>>>>>
>>>>> However, for some of my networks, I get negative values, or even NaN.
>>>>> Here
>>>>> are two examples, under the graphml format:
>>>>> http://dx.doi.org/10.6084/m9.figshare.1540708
>>>>> - scale=32.graphml
>>>>> - scale=41.graphml
>>>>>
>>>>> For the first one, the first values returned by "edge.betweenness" are:
>>>>>    [1] 1904887544.08 1904887544.08 1896303182.39 1951787568.72
>>>>> 1203043060.76
>>>>>    [6] 1270869072.68  622780616.09  667964773.27  279064394.68
>>>>> 309184936.21
>>>>>   [11]  135403467.81  155266075.94   51600202.02   60120695.31
>>>>> 21113003.39
>>>>>   [16]   24783603.89    6275147.30    6937885.52    1347425.01
>>>>> 1002544.99
>>>>>   [21]     150574.42    -327097.77    -711849.38   -1430744.36
>>>>> -246214.20
>>>>>   [26]    -602827.15    -230344.97    -484630.12    -297768.08
>>>>> -492364.06
>>>>>
>>>>> For the second one, all the returned values are NaN.
>>>>>
>>>>> Note that all these weights are positive by definition. They even are
>>>>> non-zero since no two nodes hold the same position, by construction. I
>>>>> also
>>>>> checked this programmatically. Moreover, there are no multiple links,
>>>>> also
>>>>> by construction (and I checked with "has.multiple").
>>>>>
>>>>> I was wondering if the negative or NaN values I get are due to me
>>>>> misusing
>>>>> the function, or if this is a bug in igraph.
>>>>>
>>>>> Thanks,
>>>>> Vincent Labatut
>>>>>
>>>>> _______________________________________________
>>>>> igraph-help mailing list
>>>>> address@hidden
>>>>> https://lists.nongnu.org/mailman/listinfo/igraph-help
>>>>>
>>>
>>>
>>> On Sat, Sep 12, 2015 at 3:08 PM, Vincent Labatut <address@hidden>
>>> wrote:
>>>>
>>>> Hello,
>>>>
>>>> I am processing the edge-betweenness of various networks using R igraph
>>>> version 7.1. Those are spatial networks (each node has a (x,y) position) 
>>>> and
>>>> I am using the "weight" option of the "edge.betweenness" function to take
>>>> the spatial distances into account. This spatial distance is stored in an
>>>> edge attribute called "dist".
>>>>
>>>> Here is the command I use:
>>>> edge.betweenness(graph=g, weights=E(g)$dist)
>>>>
>>>> However, for some of my networks, I get negative values, or even NaN. Here
>>>> are two examples, under the graphml format:
>>>> http://dx.doi.org/10.6084/m9.figshare.1540708
>>>> - scale=32.graphml
>>>> - scale=41.graphml
>>>>
>>>> For the first one, the first values returned by "edge.betweenness" are:
>>>>    [1] 1904887544.08 1904887544.08 1896303182.39 1951787568.72
>>>> 1203043060.76
>>>>    [6] 1270869072.68  622780616.09  667964773.27  279064394.68
>>>> 309184936.21
>>>>   [11]  135403467.81  155266075.94   51600202.02   60120695.31
>>>> 21113003.39
>>>>   [16]   24783603.89    6275147.30    6937885.52    1347425.01
>>>> 1002544.99
>>>>   [21]     150574.42    -327097.77    -711849.38   -1430744.36
>>>> -246214.20
>>>>   [26]    -602827.15    -230344.97    -484630.12    -297768.08
>>>> -492364.06
>>>>
>>>> For the second one, all the returned values are NaN.
>>>>
>>>> Note that all these weights are positive by definition. They even are
>>>> non-zero since no two nodes hold the same position, by construction. I also
>>>> checked this programmatically. Moreover, there are no multiple links, also
>>>> by construction (and I checked with "has.multiple").
>>>>
>>>> I was wondering if the negative or NaN values I get are due to me misusing
>>>> the function, or if this is a bug in igraph.
>>>>
>>>> Thanks,
>>>> Vincent Labatut
>>>
>>>
>>>
>>> _______________________________________________
>>> igraph-help mailing list
>>> address@hidden
>>> https://lists.nongnu.org/mailman/listinfo/igraph-help
>>>
>>
>> _______________________________________________
>> igraph-help mailing list
>> address@hidden
>> https://lists.nongnu.org/mailman/listinfo/igraph-help