Engineering Virality with Bayesian Optimization

The Core Question

Most AI systems ask:

Can we predict what will go viral?

Prediction assumes trends are passive - something we observe after they emerge.

But trends are not passive.

They are shaped by decisions:

• who sees content first
• when it is posted
• how it is amplified
• where visibility expands

So the real question becomes:

Can we optimize the conditions under which trends emerge?

This is where Bayesian Optimization (BO) enters.

---

The Big Idea

Instead of treating virality as a prediction problem, we treat it as a black-box optimization problem.

Traditional pipeline:

Data → Model → Prediction

New pipeline:

Decision → Propagation → Feedback → Optimization → Better Decision

The shift is subtle - but profound.

We move from observing trends to engineering them.

---

Trends as Propagation Systems

A trend doesn’t appear instantly. It propagates through stages.

1. Engagement Initiation

Early signals emerge:

• likes
• comments
• shares
• watch time

These become a feature vector:

‘x ∈ ℝ^d’

---

2. Amplification

Influencers and network structure drive spread:

• resharing velocity
• high-degree nodes
• algorithmic boosts

Amplification introduces uncertainty - exactly what BO exploits.

---

3. Visibility Expansion

Algorithms increase exposure:

• explore pages
• trending feeds
• recommendations

This stage determines whether content dies or explodes.

---

4. Virality

Sustained exponential spread occurs across the network.

This becomes the optimization target.

---

The Optimization Problem

Instead of predicting virality directly:

‘max_{x ∈ X} f(x)’

Where:

• x = trend seeding strategy
• f(x) = influence spread (unknown)

We cannot write this function analytically.

We can only evaluate it.

That makes it a perfect black-box optimization problem.

---

Influence Spread as the Objective

We evaluate f(x) using the Independent Cascade (IC) model.

Independent Cascade Simulation

1. Seed nodes activate.
2. Activated nodes attempt to activate neighbors with probability (p_{uv}).
3. Process repeats until no new activations occur.

Expected influence:

‘f(x) = E_IC[Reach(S_x)]’

Estimated using Monte Carlo simulations.

Each evaluation is expensive - which is exactly where BO shines.

---

Bayesian Optimization Core

Surrogate Model

‘f(x) ~ GP(μ(x), k(x, x’))’

Gaussian Processes model:

• predicted influence
• uncertainty over unknown regions

---

Acquisition Function

‘a(x) = μ(x) + κ σ(x)’

This balances:

• exploitation (high-performing strategies)
• exploration (uncertain but promising strategies)

---

Sequential Loop

1. Evaluate initial strategies
2. Train GP surrogate
3. Optimize acquisition function
4. Test new strategy
5. Update model
6. Repeat

Optimization becomes intelligent exploration.

---

Visualizing the Process

IC Spread Simulation

flowchart LR
  %% Time-layered IC spread (t0 -> t1 -> t2 -> t3)

  subgraph T0["t0 - seeds"]
    direction TB
    S1((S1)):::seed
    S2((S2)):::seed
  end

  subgraph T1["t1 - newly activated"]
    direction TB
    A1((A1)):::active
    A2((A2)):::active
    A3((A3)):::active
  end

  subgraph T2["t2 - newly activated"]
    direction TB
    B1((B1)):::active
    B2((B2)):::active
    B3((B3)):::active
    B4((B4)):::active
  end

  subgraph T3["t3 - newly activated"]
    direction TB
    C1((C1)):::active
    C2((C2)):::active
  end

  %% Propagation edges with probabilities (example p_uv)
  S1 -- "p=0.35" --> A1
  S1 -- "p=0.22" --> A2
  S2 -- "p=0.40" --> A2
  S2 -- "p=0.18" --> A3

  A1 -- "p=0.30" --> B1
  A1 -- "p=0.12" --> B2
  A2 -- "p=0.28" --> B2
  A2 -- "p=0.20" --> B3
  A3 -- "p=0.25" --> B4

  B2 -- "p=0.15" --> C1
  B3 -- "p=0.33" --> C1
  B4 -- "p=0.19" --> C2

  %% Styling
  classDef seed fill:#ff6b6b,stroke:#111,color:#fff,stroke-width:1px;
  classDef active fill:#4dabf7,stroke:#111,color:#fff,stroke-width:1px;

---

Acquisition Function Evolution

How BO decisions evolve over iterations:

• uncertainty decreases
• exploration shifts toward promising regions

---

Search Space Trajectory

flowchart TB
  %% "Heatmap-style" BO convergence (Mermaid can't render true heatmaps,
  %% so we approximate with a colored grid + a trajectory overlay)

  subgraph H["Search Space Heatmap (approx) - BO converges toward higher values"]
    direction TB

    %% Row 1 (low to medium)
    subgraph R1[" "]
      direction LR
      A11[" "]:::v1
      A12[" "]:::v2
      A13[" "]:::v2
      A14[" "]:::v3
      A15[" "]:::v3
    end

    %% Row 2
    subgraph R2[" "]
      direction LR
      A21[" "]:::v2
      A22[" "]:::v2
      A23[" "]:::v3
      A24[" "]:::v4
      A25[" "]:::v4
    end

    %% Row 3
    subgraph R3[" "]
      direction LR
      A31[" "]:::v2
      A32[" "]:::v3
      A33[" "]:::v4
      A34[" "]:::v5
      A35[" "]:::v6
    end

    %% Row 4
    subgraph R4[" "]
      direction LR
      A41[" "]:::v3
      A42[" "]:::v4
      A43[" "]:::v5
      A44[" "]:::v6
      A45[" "]:::v7
    end

    %% Row 5 (high region / optimum neighborhood)
    subgraph R5[" "]
      direction LR
      A51[" "]:::v3
      A52[" "]:::v5
      A53[" "]:::v6
      A54[" "]:::v7
      A55[" "]:::v8
    end
  end

  %% BO evaluation trajectory overlay (sample points over iterations)
  A11 -->|t1| A22 -->|t2| A33 -->|t3| A44 -->|t4| A55

  %% Color scale (cool -> hot)
  classDef v1 fill:#e8f4ff,stroke:#9dbfe6,stroke-width:1px;
  classDef v2 fill:#cfe8ff,stroke:#9dbfe6,stroke-width:1px;
  classDef v3 fill:#a8d2ff,stroke:#7aaee6,stroke-width:1px;
  classDef v4 fill:#7ab6ff,stroke:#5a95de,stroke-width:1px;
  classDef v5 fill:#4b97ff,stroke:#2f6fce,stroke-width:1px,color:#fff;
  classDef v6 fill:#2d7be6,stroke:#1f5fb8,stroke-width:1px,color:#fff;
  classDef v7 fill:#1f5fb8,stroke:#164787,stroke-width:1px,color:#fff;
  classDef v8 fill:#163a66,stroke:#0f2745,stroke-width:1px,color:#fff;

---

Performance vs. Random Seeding (Why BO Wins)

A simple baseline for trend seeding is random selection:

• pick seed users randomly (or randomly choose strategy knobs like time/hashtags/influencer tier)
• run the propagation simulation (IC / Monte Carlo)
• record reach / engagement / virality

This is appealing because it is easy and unbiased - but it is extremely inefficient in large combinatorial spaces.

Bayesian Optimization (BO) improves over random by learning from every evaluation and deciding where to test next using a surrogate model + acquisition function. In high-noise propagation problems, this often yields stronger performance under a fixed evaluation budget.

---

What we compare

Assume the objective is expected reach:

‘f(x) = E_IC[Reach(S_x)]’

We compare two methods under the same budget (N):

• Random: sample (x1, x2,….., xN) uniformly from the strategy space
• BO: adaptively choose (x_{t+1} = \arg\max a(x)) after each evaluation

---

Metrics used for comparison

1. Best-So-Far Influence (Simple & Strong)
   Measures progress as optimization proceeds:

‘BestSoFar(t) = max_{i ≤ t} f(x_i)’

BO should rise faster and reach higher values with fewer evaluations.

1. Area Under the Best-So-Far Curve (AUBC)
   A single-number summary of sample efficiency:

‘AUBC = Σ_{t=1}^{N} BestSoFar(t)’

Higher AUBC means you achieved strong performance earlier and consistently.

1. Regret (Optimization Lens)
   If we assume the true optimum is (f(x^*)), the simple regret is:

‘r_N = f(x*) − max_{t ≤ N} f(x_t)’

BO should yield lower regret under equal budgets.

1. Stability Under Noise (Propagation Variance)
   Since MCST is stochastic, compare variance:

‘Var(f(x)) ≈ (1 / (K − 1)) Σ_{k=1}^{K} ( Reach_k(S_x) − Ȓeach(S_x) )²’

BO can incorporate uncertainty and avoid being misled by noisy single samples.

---

1) Convergence: Best-So-Far Reach vs Iterations

This diagram shows the typical qualitative behavior: BO climbs faster than random.

graph LR

subgraph Iterations
I1["Iter 1"] --> I2["Iter 2"] --> I3["Iter 3"] --> I4["Iter 4"] --> I5["Iter 5"]
end

subgraph Random
R1["10"] --> R2["18"] --> R3["22"] --> R4["28"] --> R5["31"]
end

subgraph BO
B1["10"] --> B2["25"] --> B3["40"] --> B4["55"] --> B5["65"]
end

I1 --- R1
I2 --- R2
I3 --- R3
I4 --- R4
I5 --- R5

I1 --- B1
I2 --- B2
I3 --- B3
I4 --- B4
I5 --- B5

---

2) Budget Efficiency: Same Budget, Different Outcome

With a fixed evaluation budget (e.g., 20 simulations), BO typically discovers a higher-performing strategy.

flowchart LR
  R["Random\nBest Reach ≈ 55"]:::low --> C["Fixed Budget (N evals)"]
  B["BO\nBest Reach ≈ 85"]:::high --> C

  classDef low fill:#ffd6d6,stroke:#111,color:#111;
  classDef high fill:#d6f5d6,stroke:#111,color:#111;

---

3) Why Random Fails: Search Space Explosion

Random sampling wastes many evaluations in low-performing regions.

flowchart LR
    A["Huge Strategy Space X"] --> B["Random picks mostly low/medium zones"]
    A --> C["BO builds surrogate f̂(x) + uncertainty"]
    C --> D["Acquisition picks high-potential regions"]
    D --> E["Faster discovery of strong seeding strategies"]

---

Under the same evaluation budget (N), • BO should achieve higher BestSoFar reach • higher AUBC • lower simple regret • and often more stable improvement under Monte Carlo noise.

A realistic target framing for synthetic IC experiments: • BO improves best reach by ~20–50% vs random under tight budgets (e.g., 20–50 evaluations), depending on: • network structure • seed set size • noise level (number of MC rollouts per evaluation) • dimensionality of strategy space

Performance vs Random Seeding (Figures)

Best-so-far convergence (mean ± 95% CI)

Fixed-budget outcome (bar chart)

Distribution of best reach (violin plot)

---

Real-Graph Research Artifact: Higgs Retweet Network

The synthetic IC figures above explain the optimization intuition. The research-artifact version uses the SNAP Higgs retweet graph: 223,833 nodes, 308,596 edges, and a seed budget of k = 20.

Real data: BO reaches the high-spread region in fewer evaluations than random search, while top-degree remains an extremely strong baseline.

The actual lesson is structural: the search-space design decides whether BO refines a strong heuristic, collapses, or finds a better diverse subset.

Diversity only pays when it is constrained to nodes that can actually carry a cascade.

For the full build log and limitations, see Influence Maximization using BO.

---

Why This Is Innovative

Bayesian Optimization is commonly used for:

• protein synthesis
• material discovery
• hyperparameter tuning

This work applies BO to:

Social influence optimization

Key innovation:

• trends become optimization landscapes
• virality becomes an objective function
• strategy becomes a decision variable

---

Why Not Reinforcement Learning?

RL works best when:

• dense rewards exist
• continuous interaction is available

Here:

• evaluations are expensive
• rewards are delayed
• sample efficiency matters

BO provides better uncertainty-aware exploration.

RL becomes a future extension.

---

Challenges & Limitations

No idea is perfect.

Key challenges include:

• high evaluation noise from Monte Carlo simulations
• local traps in optimization space
• high-dimensional strategy vectors
• GP scalability limitations

These motivate future research directions.

Notes / Caveats • BO can be misled by noise if (K) is too small (not enough MC rollouts). • BO may fall into local traps if exploration is too low. • Random can look competitive in extremely high noise or poorly specified search spaces.

So comparisons will include: • confidence intervals over multiple runs (repeat experiments with different random seeds) • optionally varying (K) (MC rollouts) to show robustness.

---

The Deeper Insight

Traditional view

BO optimizes vectors.

New view

BO navigates complex social decision spaces.

This reframes social systems as:

• structured optimization landscapes
• sequential experimentation environments

---

The Mathematical Engine Behind Trend Optimization

At the heart of this system is a precise mathematical framework that turns social propagation into an optimization problem. Below is a conceptual breakdown of each core equation and how it connects to virality engineering.

---

1. Estimating Influence Spread

‘E_IC[Reach(S_x)] ≈ (1 / K) Σ_{k=1}^{K} Reach_k(S_x)’

This equation approximates the expected influence spread under the Independent Cascade (IC) model.

• (S_x) = seed set selected by strategy (x)
• {Reach}_k(S_x)) = number of activated nodes in simulation (k)
• (K) = number of Monte Carlo simulations

Because influence propagation is stochastic, we simulate it multiple times and average the results. This gives us a stable estimate of expected virality.

Each evaluation is computationally expensive - making Bayesian Optimization ideal.

---

2. Constructing the Virality Score

‘ViralityScore(x) = Σ_i w_i · f_i(x)’

Virality is not a single signal - it is a weighted combination of measurable engagement features:

• (f_i(x)) = engagement feature (e.g., share velocity, clustering coefficient, watch-time retention)
• (w_i) = learned importance weight

This linear aggregation transforms raw social signals into a structured objective function.

---

3. Learning Optimal Feature Weights

‘L(w) = - Σ_{t=1}^{T} ( y_t - Σ_i w_i f_i(x_t) )²’

This loss function learns weights (w_i) that best align predicted virality with observed viral outcomes.

• (y_t) = observed virality outcome at time (t)
• (x_t) = strategy used
• (T) = number of training instances

The negative squared error formulation allows us to optimize weights using Bayesian Optimization itself, refining how virality is computed.

---

4. Acquisition Functions (Decision Rule)

Bayesian Optimization does not optimize the true function directly. It optimizes a surrogate model and uses acquisition functions to choose the next evaluation.

Expected Improvement (EI)

‘a_EI(x) = E[ max(0, f(x) - f(x⁺)) ]’

• f(x⁺) = best observed value so far

EI chooses strategies expected to outperform the current best - balancing exploration and exploitation.

---

Upper Confidence Bound (UCB)

‘a_UCB(x) = μ(x) + κ σ(x)’

• μ(x) = predicted mean
• σ(x) = uncertainty
• κ = exploration parameter

UCB explicitly trades off:

• Exploitation (high predicted influence)
• Exploration (high uncertainty)

---

5. Strategy Selection

‘x* = argmax_{x ∈ X} a(x)’

The next strategy tested is the one that maximizes the acquisition function.

This is the core decision-making step in Bayesian Optimization.

---

6. Gaussian Process Kernel Functions

The surrogate model uses kernels to measure similarity between strategies.

Radial Basis Function (RBF)

‘k_RBF(x, x’) = exp( -

x - x’

² / (2ℓ²) )’

• Smooth, infinitely differentiable
• Assumes gradual changes in influence landscape
• Controlled by length-scale (\ell)

---

Matérn Kernel

‘k_Matern(x, x’) = (2^(1-ν) / Γ(ν)) · ( (√(2ν)

x - x’

) / ℓ )^ν · K_ν( (√(2ν)

x - x’

) / ℓ )’

• More flexible than RBF
• Better suited for noisy, less smooth social systems
• Controlled by smoothness parameter (\nu)

---

7. Edge Activation Probability

‘p_uv’

This represents the probability that an activated node (u) activates neighbor (v) in the Independent Cascade model.

These probabilities encode:

• strength of connection
• historical interaction frequency
• platform-specific influence likelihood

---

Why This Matters

These equations collectively transform:

• engagement signals
• network structure
• uncertainty
• propagation randomness

into a sequential optimization framework.

Rather than predicting trends passively, this framework:

1. Models uncertainty over virality
2. Selects promising strategies
3. Learns from outcomes
4. Improves future decisions

The mathematics is not just descriptive - it is prescriptive.

It enables AI to navigate social systems intelligently.

---

Future Directions

Where this research can go next:

• Graph Neural Network surrogates
• RL-based adaptive strategies
• Multi-fidelity Bayesian Optimization
• Dynamic, evolving social graphs
• Real-time autonomous trend engineering

This is where sequential decision-making and BO begin to merge.

---

The Big Idea - Final Thought

Most systems try to predict trends after they happen.

This work asks something different:

What if AI could learn how to create them?

By combining influence propagation models with Bayesian Optimization, we move from passive forecasting toward active, uncertainty-aware decision making.

Sometimes optimization isn’t about finding the answer.

It’s about learning how to move through the space.

---

References

• Kempe et al. (2003). Maximizing the Spread of Influence through a Social Network
• Snoek et al. (2012). Practical Bayesian Optimization
• García-Hernández et al. (2022). Bayesian Optimization for Influence Maximization
• Liang (2024). Bayesian Optimization of Functions over Node Subsets in Graphs

---

Author Note

Written as part of my exploration in AI for Sequential Decision Making
MS Data Science - University of Minnesota