Multi-Objective Ranking

Why Single-Objective Optimization Fails

Early recommendation systems optimized a single metric: Click-Through Rate (pCTR).

This resulted in Clickbait Proliferation:

• The model learned to recommend sensational titles ("You won't believe what happened next!").
• Users clicked, realized the video was garbage, and closed the app after 3 seconds (High Bounce Rate).

  SINGLE-OBJECTIVE CTR FAILURE:
  High Clicks (pCTR = 95%) + Low Watch Time (3 seconds) + High Dislikes ──► USER DISSATISFACTION!

To build sustainable products, modern platforms (YouTube, TikTok, Amazon) use Multi-Objective Ranking to optimize multiple engagement signals simultaneously:

┌─────────────────────────────────────────────────────────────┐
│ MULTI-OBJECTIVE ENGAGEMENT SIGNALS                          │
│ 1. P(Click):       Probability user clicks impression.     │
│ 2. E[WatchTime]:   Expected completion watch time in sec.  │
│ 3. P(Like/Share):  Probability user actively engages.      │
│ 4. P(Dislike/Hide):Probability user expresses dissatisfaction.│
└─────────────────────────────────────────────────────────────┘

Architecture 1: Multi-Gate Mixture-of-Experts (MMoE - Google, 2018)

Training 5 independent neural networks for 5 objectives requires massive GPU compute.

MMoE (Ma et al., 2018 / Google) shares lower-level Expert Sub-Networks across all tasks while maintaining Task-Specific Gating Networks:

                       MMoE ARCHITECTURE

  [ Task 1 Head: pClick ]      [ Task 2 Head: WatchTime ]     [ Task 3 Head: pShare ]
            ▲                              ▲                             ▲
  [ Gating Net 1 ]               [ Gating Net 2 ]              [ Gating Net 3 ]
            └──────────────────────┼─────────────────────────────┘
                                   ▼
  ┌─────────────────────────────────────────────────────────────┐
  │ SHARED EXPERT SUB-NETWORKS (Expert 1, Expert 2, Expert 3)   │
  └──────────────────────────┬──────────────────────────────────┘
                             ▲
                    Input Features x

• Shared Experts: Learn universal representations (user history, video topic features).
• Task Gates: Learn which specific experts to weight for each individual task objective.
• Advantage: Solves Negative Transfer (where optimizing Task A harms Task B performance).

Architecture 2: The Value Function (Scalarization)

How do you convert 4 predicted probabilities into a single scalar ranking score to order items in a user feed?

Use a Production Value Function:

$$\text{Final Score} = w_{\text{click}} \cdot \hat{p}{\text{click}} + w{\text{watch}} \cdot f(\hat{\text{WatchTime}}) + w_{\text{share}} \cdot \hat{p}{\text{share}} - w{\text{hide}} \cdot \hat{p}_{\text{hide}}$$

┌──────────────────────────┬──────────────────────────┐
│ BUSINESS WEIGHT          │ PRODUCT IMPACT           │
├──────────────────────────┼──────────────────────────┤
│ Increase w_watch         │ Increases long videos &  │
│                          │ deep user engagement.    │
│ Increase w_share         │ Promotes viral, high-    │
│                          │ quality content.         │
│ Increase w_hide          │ Strictly filters out     │
│                          │ clickbait & spam.        │
└──────────────────────────┴──────────────────────────┘

Product managers tune hyperparameter weights $w_1, w_2, w_3$ via live A/B testing to balance short-term clicks against long-term user retention.

Say this out loud

Multi Objective Ranking optimizes recommendations across competing signals like clicks, watch time, shares, and hides. Google MMoE uses shared expert sub networks with task specific gating networks to predict multiple objectives without negative transfer. Predicted probabilities are combined into a final ranking score using a scalarized value function with business hyperparameter weights.

Followups to expect

1. What is Pareto Efficiency in Multi-Objective Ranking? A state where no single objective (e.g. Watch Time) can be improved without degrading another objective (e.g. Ad Revenue), mapped as a Pareto Frontier during weight tuning.
2. How do you handle zero-inflated targets like Purchases or Shares? Using a Two-Stage hurdle model: first predict binary probability $P(\text{Share} > 0)$, then predict conditional magnitude $E[\text{Shares} \mid \text{Share} > 0]$.