Neural Network Media Mix Modeling:

An Integrated Mediation Analysis Framework

Introduction: Rethinking Media Mix Models Through Mediation Analysis

Media mix modeling traditionally attempts to directly map marketing spend to revenue, often treating the process as a black box. However, we know that marketing operates through a chain of effects: spend generates impressions, and these impressions then influence consumer behavior to drive revenue. This indirect pathway represents a classic mediation relationship, where impressions serve as the mediator between marketing investment and financial outcomes.

Traditional mediation analysis typically involves separate models for each relationship:

1. Modeling spend's effect on impressions
2. Modeling impressions' effect on revenue
3. Combining these effects to understand total impact

The innovation lies in using a multi-stage neural network to integrate these traditionally separate steps into a single, cohesive framework. This unified approach allows the model to learn the entire causal chain simultaneously, capturing complex interactions and dependencies that might be missed when these relationships are modeled separately.

The Multi-Stage Neural Network: A Unified Framework

The architecture explicitly encodes the mediation structure through its stage design:

Stage 1: Marketing Spend → Impressions

The first stage uses dense neural network layers to capture how different marketing channels generate impressions. This mirrors the first step of traditional mediation analysis but allows for non-linear relationships and interaction effects between channels.

Stage 2: Impressions → Revenue with Time Consideration

A critical insight in marketing is that the relationship between impressions and revenue isn't instantaneous. Different channels may have varying lag effects - for example, a display ad might influence purchase decisions weeks after viewing, while a search ad might have more immediate impact. We use LSTM (Long Short-Term Memory) layers to capture these temporal dynamics, allowing the model to:

• Learn channel-specific lag patterns
• Account for cumulative exposure effects
• Adapt to varying response times across different markets and seasons

Stage 3: Integrated Effects

The final stage combines both direct and indirect pathways, allowing the model to capture:

• Mediated effects (spend → impressions → revenue)
• Direct effects (spend → revenue)
• Cross-channel interactions
• Geographic variations

This integrated approach offers several advantages over traditional separated mediation analysis:

• Joint optimization of all parameters
• Consistent treatment of uncertainty
• Ability to capture complex interactions between stages
• More efficient use of data

Implementation and Temporal Dynamics

The implementation acknowledges two key realities of marketing:

1. Time-Lagged Effects: Revenue often materializes days or weeks after marketing activities. Our LSTM layers explicitly model these temporal dynamics, learning:

   • Optimal attribution windows for each channel
   • Interaction effects over time
   • Seasonal patterns in response rates

2. Geographic Heterogeneity: Marketing effectiveness varies by region. Our model incorporates this through:

   • Regional dummy variables
   • Market-specific response patterns
   • Local temporal effects

Technical Framework

The model is implemented using TensorFlow/Keras, leveraging modern deep learning capabilities to handle the complexity of integrated mediation analysis. Key technical features include:

# Simplified architecture overview
spend_input = Input(shape=(sequence_length, n_features))

# Stage 1: Spend → Impressions
impressions = Dense(units)(spend_input)

# Stage 2: Time-Aware Impression → Revenue
temporal_effects = LSTM(units)(impressions)

# Stage 3: Combined Effects
combined = Concatenate()([spend_input, temporal_effects])
revenue = Dense(1)(combined)

Business Impact

This integrated approach provides several advantages:

1. More Accurate Attribution: By modeling the entire causal chain simultaneously, we achieve more accurate attribution of revenue to marketing activities.

2. Temporal Understanding: The LSTM components reveal how different channels contribute to revenue over time, enabling better planning of marketing activities.

3. Channel Interactions: The unified framework captures how different marketing channels work together, rather than treating them independently.

4. Actionable Insights: The model provides clear guidance for:

   • Budget allocation across channels
   • Timing of marketing activities
   • Geographic targeting
   • Campaign optimization

Future Directions

This framework opens several exciting avenues for future development:

• Incorporation of competitor activities
• Addition of more sophisticated market response curves
• Integration of customer journey data
• Extension to handle multi-product scenarios

---

Requirements

tensorflow>=2.0.0
pandas
numpy
scikit-learn