On personal recommendations in social networks

I. INTRODUCTION

• Automated recommendations influence communities in domains including transportation, energy systems, and social networks.Personalized recommendations also appear in political discourse and dietary suggestions.Source: §I · p. 1
• Existing recommendation models typically omit how individuals psychologically process recommendations, despite this affecting the analysis and design of social-network algorithms.The proposed model draws on confirmation bias, while replacing discontinuous bounded-confidence rules with a smooth influence function.Source: §I · p. 1
• The model extends DeGroot consensus dynamics with an analytically tractable Gaussian influence function consistent with confirmation bias.The paper examines how the recommendation value shapes broadcast behavior, including a fold bifurcation and additional equilibria above a critical threshold.Source: §I · p. 1
• For time-varying targeted recommendations, a greedy policy is optimal for reaching the desired state in the shortest time and guarantees convergence.The paper extends convergence guarantees to uncertain agent sensitivity using a worst-case estimate.Source: §I · p. 2

II. MODEL AND PROBLEM FORMULATION

• The system models n agents’ opinions on a connected undirected graph, with opinions and recommendations bounded in [−1, 1].The state x(t) records opinions, and u(t) records recommendations.Source: §II · p. 2
• Each update combines neighbor influence through W with personal recommendation influence through ψ, weighted by the trade-off parameter a.At a = 1, the model reduces to the DeGroot model.Source: §II · p. 2
• The Gaussian acceptance function treats a recommendation matching the current opinion as neutral and decreases smoothly as their distance grows.Larger sensitivity α narrows the acceptance window by making the influence function decay faster; the analysis assumes homogeneous sensitivities.Source: §II · p. 2
• The problem asks whether broadcast recommendations can reach target x_d = 1, which targeted recommendations guarantee convergence, and whether convergence holds for uncertain sensitivity.Sensitivity uncertainty is specified as α ∈ [α−, α+].Source: §II · p. 2

III. TIME-INVARIANT BROADCAST RECOMMENDATIONS

• Under homogeneous recommendations, the consensus space is invariant, so consensus equilibria can be identified from the scalar fixed-point condition ξ=ψ(ξ,u).The analysis varies the constant broadcast value u as the bifurcation parameter.Source: §III · pp. 2–3
• There can be at most three consensus equilibria, each component lies in [0,u], but the authors leave the absence of disagreement equilibria as a conjecture.The bound on the number of consensus equilibria is established; the conjecture concerns non-consensus equilibria.Source: §III · p. 3
• At α=4, u*=0.7835; below it there is one consensus equilibrium, while crossing it creates two more in a fold bifurcation.The new equilibria appear when u crosses the threshold.Source: §III · p. 3, Figure 1 · p. 4
• Above u*, x*=u1 and x*s are locally asymptotically stable, while x*us is unstable.For u<ũ*, the sufficient contraction condition gives a unique globally exponentially stable equilibrium x*=u1.Source: §III · p. 4
• For initial conditions near the target, increasing u brings the equilibrium x*=u1 closer to it; for general initial conditions, high broadcasts can yield worse outcomes than low ones.The distance to the target grows with the recommendation for the alternative consensus equilibrium x*s.Source: §III · p. 5

IV. TIME-VARYING TARGETED RECOMMENDATIONS

• The section designs time-varying targeted recommendations after analyzing the limitations of constant broadcast policies.The greedy policy is derived as globally optimal in Section IV-A.Source: §IV · p. 5

IV-A

• A greedy policy is derived and shown to be globally optimal.The paper describes it as reaching the desired state in the shortest time.Source: §IV · p. 5

IV-B

• The analysis extends to a robust policy for uncertain sensitivity α.This extension is addressed in Section IV-B.Source: §IV · p. 5

A. Greedy targeting

• At each step, the controller chooses personal recommendations to maximize influence and move agents closer to x_d = 1.The pointwise optimal input maximizes each agent’s influence response.Source: A. Greedy targeting · p. 5
• The optimal recommendation’s influence response is strictly increasing in the agent’s opinion.The unconstrained optimum is clipped to the allowed recommendation set when its constraint activates.Source: A. Greedy targeting · p. 5
• No input-cost penalty is included because recommendations are interpreted as content choices rather than costly control effort.The paper gives showing content in a recommender system as an example.Source: A. Greedy targeting · p. 5
• The greedy policy is globally optimal for every componentwise-monotone terminal objective.Its trajectory dominates other trajectories component-wise at each time.Source: A. Greedy targeting · p. 5
• The greedy policy converges to the desired state x_d = 1 from any initial condition.The proof establishes that the minimum opinion increases toward 1.Source: A. Greedy targeting · pp. 5–6
• The policy depends only on each agent’s state and sensitivity, so it can be implemented by distributed actors without full system knowledge.The actors need to communicate the global sensitivity parameter.Source: A. Greedy targeting · p. 6

B. Robust greedy policy

• The policy designed with α̂ = α+ converges to x_d = 1 from any initial state for every α_true ∈ [α−, α+].The upper-bound estimate assumes the narrowest recommendation acceptance window.Source: B. Robust greedy policy · p. 6
• When α_true ≪ α+, the robust policy converges more slowly because it is more conservative than necessary.The controller may otherwise recommend too far from an agent’s current opinion.Source: B. Robust greedy policy · p. 6

V. NUMERICAL EXAMPLE

• The example uses 20 agents on a connected random graph, with a = 0.5 and initial opinions drawn uniformly from X^n.Each edge exists with probability 0.2, and the edge weights are drawn uniformly and normalized.Source: §V · p. 6
• Large broadcast recommendations can produce different equilibria by initial condition, and some trajectories move opposite the recommendation.The example contrasts u = 0.5 with u = 0.9 and notes that initial opinions must be sufficiently biased for broadcast convergence.Source: §V · pp. 6–7
• For uncertain sensitivity α, the robust greedy policy with α̂ = α+ converges more slowly than perfect estimation α̂ = α_true; choosing α̂ = α− can prevent convergence to xd.Source: §V · p. 7

VI. CONCLUSION

• The proposed model uses a smooth Gaussian influence function consistent with behavioral-psychology effects and is analytically tractable.The paper presents it as a foundation for analyzing personal recommendations.Source: §VI · p. 7
• The broadcast analysis finds that the desired target is not always reachable, depending on initial opinions.The paper also designs optimal targeted recommendations and gives robust convergence results for uncertain sensitivities.Source: §VI · p. 7
• Future work will consider heterogeneous sensitivities and social-term weights, as well as opinion processes that capture biases and stubbornness.Source: §VI · p. 7

• The appendix states that it collects the first and second derivatives of ψᵢ(xᵢ, uᵢ).Source: Appendix I · p. 7
• The Lipschitz constant is given by Lψ(uᵢ) = 2αe^(-1/2)uᵢ.Source: Appendix I · p. 7
• For fixed uᵢ = u*, ψᵢ(xᵢ, u*) reaches its maximum value u* at xᵢ = u*, where its derivative with respect to xᵢ is zero.Source: Appendix I · p. 7