Subject: Re: [xsl] RDF graph to SVG force-directed layout From: "Martynas Jusevičius martynas@xxxxxxxxxxxxx" <xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx> Date: Fri, 26 Mar 2021 22:22:33 -0000 |
Hi Michael, I wanted to pick up this thread as I started working on this stylesheet again. I've implemented some of the optimizations suggested in this thread. https://github.com/AtomGraph/Web-Client/blob/master/src/main/webapp/static/co m/atomgraph/client/xsl/converters/RDFXML2SVG.xsl I want to apply the (node, x, y, list-of-adjacent-nodes) tuple strategy you suggested. But what would be an efficient data structure for it in practice? Something like the following? <xsl:for-each select="svg:g"> <xsl:variable name="tuple-set" as="map(xs:string, item())"> <xsl:map> <xsl:map-entry key="'node'" select="."/> <xsl:map-entry key="'x'" select="svg:circle/@cx"/> <xsl:map-entry key="'y'" select="svg:circle/@cy"/> <xsl:map-entry key="'adjacent-nodes'" select="/svg:svg/svg:g[svg:g[@class = 'property']/@resource = current()/@about]/svg:g[@class = ('subject', 'object')]"/> <xsl:map-entry key="'non-adjacent-nodes'" select="..."/> </xsl:map> </xsl:variable> </xsl:for-each> Another thing I wasn't sure how to fix was your remark: <xsl:param name="force-nodes" select="key('nodes', ('circle', 'rect'))" as="element()*"/> where the key is <xsl:key name="nodes" match="svg:g[svg:circle] | svg:g[svg:rect]" use="concat(svg:circle/local-name(), svg:rect/local-name())"/> This is going to create a key with a very small number of entries, each indexing a large number of nodes, and that doesn't feel like an efficient thing to do. On Thu, Oct 15, 2020 at 11:51 AM Michael Kay mike@xxxxxxxxxxxx <xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx> wrote: > > OK, I'm starting to see where the adjacency logic is, around lines 335-350. > > It's computing some kind of "attraction force" for every pair of nodes in the graph, where the force is negative if the two nodes aren't cconnected. > > The connectivity of nodes is unchanged for each iteration, but is being recomputed on each iteration. If the data that you process on each iteration is a set of tuples comprising > > (node, x, y, list-of-adjacent-nodes) > > then you can precompute the list-of-adjacent-nodes (and possibly even list-of-non-adjacent-nodes) at the start, and only modify x and y each time round the loop. > > It would be nice to find a way of reducing the number of distant non-adjacent nodes you have to consider as the algorithm converges. Perhaps there is some threshold where if a non-adjacent node is more than a certain distance away, you can drop it from the list-of-non-adjacent-nodes that you have to consider next time around, so that over time, you're only looking at your position relative to nodes that are nearby? > > Michael Kay > Saxonica > > On 15 Oct 2020, at 10:17, Michael Kay mike@xxxxxxxxxxxx <xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx> wrote: > > Having said all this, however, making the inner loop of your algorithm 5 times faster is only going to get you from 5 minutes to 1 minute, and to do better than that (as Liam points out) you need to improve the algorithm. > > Doing a fixed number of iterations (50) seems rather crude; is there some way you could detect convergence and stop when the results are good enough? > > Might there be some way of doing initial placement that's better than just random positioning? For example, perhaps it's the case that in many datasets, the order of the input nodes reflects some kind of proximity metric, and therefore computing an initial position based on position in the input sequence would give faster convergence. > > I haven't really worked out what the criteria are for computing node proximity. I can't see any logic that decides which nodes you want to be near. > > Michael Kay > Saxonica > > On 15 Oct 2020, at 09:42, Michael Kay mike@xxxxxxxxxxxx <xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx> wrote: > > This deserves closer study than I have time to give it. > > From a quick look, my attention is drawn to the template rule on line 312. > > Firstly, the variable $net is a sequence of elements that are essentially (x, y) pairs, and my instinct is that using sequence of maps would cut a lot of node-creation cost (and a bit of number-to-string and string-to-number conversion). In fact, since you are computing the (x, y) pairs only in order to then compute the sum of the x's and the sum of the y's, it seems unneccessary to capture these values in a list in the first place; why not compute the two totals as you go by using xsl:iterate in place of xsl:for-each at line 325? > > Line 313 is > > <xsl:param name="force-nodes" select="key('nodes', ('circle', 'rect'))" as="element()*"/> > > where the key is > > <xsl:key name="nodes" match="svg:g[svg:circle] | svg:g[svg:rect]" use="concat(svg:circle/local-name(), svg:rect/local-name())"/> > > This is going to create a key with a very small number of entries, each indexing a large number of nodes, and that doesn't feel like an efficient thing to do. > > At lines 339 and 343, should the following-sibling calls be limited (using [1]) to the immediatelty following sibling? > > At line 343, should the generate-id() comparison be replaced by "is"? (The XJ compiler does this optimisation for you, XX doesn't). > > Finally, line 360-366 is a classic case of a subtree copy that's making a very small change to an existing tree; I've written several papers that attempt to address the problem that this is very inefficient because it involves copying the whole tree. > > You might consider, instead of making incremental modifications to the attributes of the svg:g elements, maintaining for each of these elements a map containing the original svg:elements and the latest computed values of the attributes that are being modified. Changing a single property in a map is much more efficient that changing a single attribute of an element. > > This would also address another issue: on each iteration where a node is repositioned, you're generating an SVG @transform attribute to reflect the new position, and then on the next iteration, you are parsing this attribute to compute the current position. Much better, surely, to maintain the actual coordinates. > > Michael Kay > Saxonica > > > On 14 Oct 2020, at 22:00, Martynas JuseviD ius martynas@xxxxxxxxxxxxx <xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx> wrote: > > Hi, > > could anyone suggest any optimizations to this stylesheet that > transforms a graph encoded as RDF/XML to an SVG directed graph layout: > https://github.com/AtomGraph/Web-Client/blob/develop/src/main/webapp/static/c om/atomgraph/client/xsl/converters/RDFXML2SVG.xsl > > Output example: https://twitter.com/namedgraph/status/1316476355874304001 > > The problem is that it's quite slow: <100 nodes and 5 steps take a few > minutes running on Saxon-JS 2 in Firefox or Chrome. > > It's based on a paper on force directed layout in XSLT: "GraphML > Transformation": > http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.182.3680&rep=rep1&ty pe=pdf#page=58 > > The algorithm: > 1. position resource nodes (optionally also literals) randomly. > (TO-DO: position on an ellipse?) > 2. move nodes in a loop using the force-directed algorithm > 3. draw lines between the nodes, calculating the correct intersection > with the node border > > Note: only "flat" RDF/XML (properties grouped into descriptions; no > nesting) is supported. It's called RDFXML_PLAIN in Jena. > > If anyone would like a sample file, I can easily provide :) > > > Martynas > atomgraph.com > > > > XSL-List info and archive > EasyUnsubscribe (by email) > > > XSL-List info and archive > EasyUnsubscribe (by email)
Current Thread |
---|
|
<- Previous | Index | Next -> |
---|---|---|
Re: [xsl] Find/replace algorithm, Dimitre Novatchev dn | Thread | Re: [xsl] RDF graph to SVG force-di, Martynas Jusevičius |
Re: [xsl] Find/replace algorithm, Albright, Jim jim_al | Date | Re: [xsl] Find/replace algorithm, Joel Kalvesmaki dire |
Month |