Contribution
It finds the reason subjective logic (Jøsang’s SL) cannot be used as a flow-based reputation metric on arbitrary graphs and fixes it. SL’s consensus adds evidence, but its discounting multiplies belief masses, so the two do not cooperate: x ⊗ (y ⊕ z) ≠ (x ⊗ y) ⊕ (x ⊗ z). The paper defines a scalar multiplication of evidence, builds discounting from it, proves the right distribution law and the absence of double-counting for transported evidence, and runs an EigenTrust-style fixed point on a BitTorrent dataset.
Key insight
Trust in a recommender should scale the evidence the recommender passes on, not reshape it. If discounting is “keep a fraction g(x) of y’s evidence”, then sending the same evidence along several paths sums fractions of one body of evidence, and the referral links serve only as weights, so nothing is counted twice.
Assumptions
- Opinions as evidence. x = (p, n, c)/(p + n + c) for positive and negative evidence p, n; Theorem 1 derives this map from four requirements, leaving c > 0 free (c = 2 in SL).
- No dogmatic opinions. u > 0 throughout (the space Ω′), so evidence is finite.
- A weighting function g : Ω′ → [0, 1], unspecified in general; g(x) = x_b, √x_b, or p(x)/θ in the experiments.
- Referral and functional trust are separate: A holds referral opinions, T functional ones, A_xx = U to stop self-rating; all opinions are published.
Key results
- SL’s failures (§3–4). Examples 2–3: with n(x) = 0 and much evidence in y, p(x ⊗ y) ≈ p(x)/4, so none of y’s evidence survives. Example 5: (x ⊗ y) ⊕ (x ⊗ z) = (0.314, 0.341, 0.345) but x ⊗ (y ⊕ z) = (0.227, 0.324, 0.449). The network of Figure 2 has no canonical form unless an edge is dropped; a loop has none.
- Scalar multiplication (Def. 11, Lemmas 2–4). α·x = (αx_b, αx_d, x_u)/(α(x_b + x_d) + x_u) multiplies evidence by α and distributes over ⊕.
- EBSL discounting (Def. 12, Theorem 2, Lemmas 5–6). x ⊡ y = g(x)·y: p(x ⊡ y) = g(x)p(y), discounting never lowers uncertainty, x ⊡ (y ⊕ z) = (x ⊡ y) ⊕ (x ⊡ z), chains commute; not associative. Theorem 3: no g makes ⊡ equal ⊗.
- Fully linear variant (Def. 13, Lemmas 8–10). g(x) = p(x)/θ is associative and left-distributive, but ignores negative evidence and needs θ above every positive evidence in the network; Lemma 10 says linear g must be a combination of p and n.
- Fixed point (§6). R = A ⊕ (offdiag R ⊡ A), F_iP = T_iP ⊕ ⊕_j R_ij ⊡ T_jP. Evidence about j in R can never exceed the evidence about j in A, so the iteration cannot run to full belief. In a loop R_12 = A_12 ⊕ (R_12 ⊡ A_23) ⊡ A_32; with A_32 near full belief, R_12 goes to full belief “even if user 1 has little trust in user 2”.
- Experiments (§7). On Figure 2’s network SL’s naïve recursion counts evidence twice, its canonical form discards it, EBSL with g = x_b or √x_b keeps the evidence ratio and the amount; ⊙ loses most of it. Tribler (10,364 nodes in 200 clusters): convergence in 47 (√x_b), 24 (x_b) and 9 (⊙) iterations, and the naïve SL recursion is still at 10⁻⁸ after 1000.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | SL’s ⊗ does not distribute over ⊕ and so double-counts in non-canonical networks | strong | Example 5, checked; Eqs. (8)–(9) |
| C2 | EBSL’s ⊡ distributes over ⊕ and transports evidence without double-counting | strong | Lemma 5, Lemma 11 |
| C3 | The EBSL fixed point handles arbitrary networks, loops included, without discarding information | moderate | the evidence bound of §6.2; convergence observed, not proved |
| C4 | EBSL gives “more accurate” reputations | weak | no ground truth; the authors say numerics cannot prove which is right |
| C5 | EBSL with suitable g is preferable to ⊗ and ⊙ | moderate | evidence propagation on synthetic and Tribler data |
Concepts
- Opinion: (belief, disbelief, uncertainty), summing to 1.
- Consensus ⊕: addition of evidence.
- Referral trust / functional trust: trust in a party’s recommendations / in its performing a function.
- Canonical form: a network expression in which each edge occurs once.
- Double counting: ⊕ applied to opinions with shared evidence.
Connections
It builds on Jøsang’s SL (2001) and his canonical-form network analysis, and on the authors’ flow-based reputation metric (Simone, Škorić and Zannone 2012), itself in the line of EigenTrust and PageRank. Cerutti et al.’s geometric discounting operators are compared and rejected. None of these is registered here.
Bearing on the record
- CLAIM-282. EBSL’s fixed point is a damped recursive delegation with self-rating forbidden (A_xx = U), the design CLAIM-282 is about, in an evidence algebra. The paper’s loop case shows the claim’s point: R_12, user 1’s trust in user 2, depends on user 3’s opinion of 2 weighted by 1’s trust in 3, which depends on R_12 itself, and a near-certain A_32 can make R_12 near-certain whatever A_12 says. What EBSL adds is a ceiling: no user’s evidence about j can exceed the evidence about j that others hold directly, so recycling cannot create evidence, only re-weight it. It does not make i’s standing independent of i’s own opinions, as one-hop allocation does (CLAIM-282). Proposed dated note below.
- CLAIM-281. SL and EBSL already index trust by role in two values, referral and functional, and per proposition; F_iP is computed per P. That is a working predecessor of role-indexed endorsement with two roles. A15’s caution holds: the algebra transports evidence, it does not stop a well-trusted referrer’s evidence from dominating.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- No ground truth and no attack model: “Experiments for assessing the robustness … against attacks like slandering, self-promotion and Sybil attacks … are left for future work.”
- Convergence is observed, not proved; oscillation is possible for tuned A and damped by adding A each step.
- Choice of g is left open, and the results depend on it.
- Slips. §8 says A ⊡ B = B for g(A) > 0, but B, full belief, is excluded from Ω′; it holds as a limit (Corollary 1). Table 1 gives (5, 5) as uncertainty 0.166 where 2/12 rounds to 0.167. Lemma 10’s Riesz–Fréchet argument needs g bounded (continuous) as well as additive, and Ω* with u = 0 allowed contains dogmatic points, so the Hilbert-space structure is assumed rather than shown.
Open questions
- Under EBSL, how much can a coalition raise a member’s R by endorsing each other, given the evidence ceiling?
- Does a g that uses negative evidence keep left-distributivity in any bounded form?