Talk:A comparison of trust metrics on the Advogato social network

Why there is no link to Category "Draft papers" at the bottom of this page? --PaoloMassa 02:22, 16 July 2007 (PDT)


 * Because we hadn't yet added Category:Draft papers to Template:Paper draft :) guaka wikitalk 02:34, 16 July 2007 (PDT)

In Advogato you need to have more trust in some people, which is like putting a hierarchy on the network. Could we say that Advogato is hierarchical and PageRank decentralized? Or maybe there are more commonly used terms for this? guaka wikitalk 03:44, 26 August 2007 (PDT)

Result from a test run of r100 on my machine at IRST: [('moletrust_tm', 0.14684450158097712, 0.93110737811668332), #paolo_mole_tm ('moletrust_tm', 0.1158451890542507, 0.93110737811668332),  #guakamoletm ('outa_tm', 0.1369661159450781, 0.98220116645589928), ('outb_tm', 0.19654936864026304, 0.92172665672599607), ('intersection_tm', 0.11949396224243668, 0.99294086559332595), (' ', 0.13322551478128344, 0.99294086559332595),     #lambda g,a,b: (avg_or_none([edges_a_tm(g,a,b), intersection_tm(g,a,b)])), ('ebay_tm', 0.15047189603126115, 0.97810535852475411)]

So apparently guakamoletm is only slightly better than intersection_tm, but intersection_tm has a coverage of >99%. Now it will be interesting to overlap guakamoletm with intersection_tm. Also note that intersection_tm is a lot faster. The advogato algorithm needs more testing, but it's even way slower than the moletrusts, which take a bit more than 3 hours on power. guaka wikitalk 07:32, 9 September 2007 (PDT)

In r101 I implemented rounded, which rounds to 0.2, 0.4, 0.6, .., but I think there must be a bug, since the numbers are exactly the same! [('ebay_tm', 0.15047189603126115, 0.97810535852475411),        rounder(ebay_tm), ('outa_tm', 0.1369661159450781, 0.98220116645589928),         rounder(outa_tm), ('moletrust_tm', 0.1158451890542507, 0.93110737811668332),    rounder(guakamoletm), ('intersection_tm', 0.11949396224243668, 0.99294086559332595), rounder(intersection_tm), ('overlap_tm', 0.11995526303684981, 0.99294086559332595),     guakamole_full_tm, ('overlap_tm', 0.11995526303684981, 0.99294086559332595)]     rounder(guakamole_full_tm), guaka wikitalk 20:26, 9 September 2007 (PDT)

Ok, small bug in rounder. Fixed in r102, which I'm running now. guaka wikitalk 20:46, 9 September 2007 (PDT)

Pagerank
Localize PR, I think it is better not, we would insert too many changes into the original algorithm.

But we need to modify PR in order to accommodate weighted relationships (actually I'm curious to see if the random walk model of pagerank is affeced by the fact you change the probability of choosing an edge at every node or not) this would not be hard.

Suppose you have 3 outgoing edges, from A to b, c and d

instead of having proabilily 0.3333 each in the transition matrix (sum prob must = 1)

they will have it rescaled based on the trust statements (but always sum = 1)

a-->b [1.0]

a-->c [0.6]

a-->d [0.8]

then the transition matrix would be something like

a-->b 1.0/(1.0+0.8+0.6) = 0.416

a-->c 0.6/(1.0+0.8+0.6) = 0.25

a-->d 0.8/(1.0+0.8+0.6) = 0.333

this means that in the random walk, when the walker will be in node A it will have a higher probability of choosing b than c and this means the final pagerank of a node (number of time the walker passed over a node) will be different.

Does it make sense?


 * I guess, but I think it's already implemented like this. We just have to check the algorithms... guaka wikitalk 11:37, 24 September 2007 (PDT)


 * The current PageRank algo as defined by Page and Brin works on the graph representing the web and so it does not deal with weighted edges, since on the Web you cannot attach a weight to a link such as "this page links to http://google.com with a weight of 0.8 in [0,1]". That's why we need to change a bit the algorithm --PaoloMassa 06:17, 25 September 2007 (PDT)

A research on the growth of MAE due to other factors.
We have done some tests on the network Kaitiaki and Advogato, to determine a law for these functions:
 * 1) how increases the MAE, with the increase of the controversiality (for the trust metric moletrust, random, intersect, PageRank, edges_a/b)
 * 2) how increases the MAE, with the increase of the horizon in moletrust trust metric
 * 3) how increases the time of computation for the moletrust trust metric with the increase of the horizon

Kaitiaki Network:
1) For the network Kaitiaki we compared the intersection trust metric with the edges_a and random trustmetric. This is the graph for the intersection trust metric:



analysing the graph seems that the MAE increases with the increase of the controversiality, as we expected. These are the graphs for the random, edges_a, edges_b and PageRank trust metrics

Random tm:



Edges A tm:



Edges B tm:



PageRank tm:



We can say that all these metrics give us approximately the same result. For the moletrust trustmetric with the best parameters (calculated empirically) this is the graph:



We can see that the function growth seems to be the same as the other trustmetrics, but the error starts from a bigger value. For this reason this trust metric is the worst.

2) Before we talked about an empirically calculation of the best moletrust parameters for the trust metric. Now we are going to explain how we obtained these results. First of all, we wrote a function that test, for each possible values of the parameters, the MAE error and save it. This function writes a graph as follows:



3) This is the graph that says how the time of computation increases with the increase of the horizon



Advogato Network
For the advogato network we have few data (for the lenght of the computation on this large network)

1) This is the MAE error for the intersection trust metric.



It seems to have a quadratic growth but we should check this more deeply.

2/3) For the moletrust trust metric this is the graph that indicates how the MAE decreases with the increase of the horizon



This is a good result, but the time of computation has the same percentage of growth.



--Ciropom 09:33, 21 April 2008 (PDT)