Advanced

Design Ride ETA Prediction

A complete walkthrough of designing an ML system that predicts how long a ride will take. This is one of the most technically challenging ML system design problems because it combines spatial data, temporal patterns, real-time signals, and strict accuracy requirements.

Step 1: Clarify Requirements

ML Problem Formulation

# Problem formulation
# Business goal:    Accurate ride duration for pricing, matching, and user trust
# ML task:          Regression (predict trip duration in seconds)
# Input:            Origin, destination, route, time, traffic, weather
# Output:           Predicted duration (seconds) + confidence interval
# Training data:    Completed trips with actual duration
# Loss function:    Huber loss (robust to outliers from unusual trips)
# Key constraint:   Under-prediction worse than over-prediction (user trust)

Step 2: High-Level Architecture

# Architecture overview
#
# [Ride Request]
#   |
# [Route Planning Service] --> Returns candidate routes (from mapping service)
#   |
# [Feature Assembly]
#   |-- Static features: road segments, speed limits, intersections
#   |-- Real-time features: current traffic, weather, events
#   |-- Historical features: typical travel time for this route/time
#   |
# [Segment-Level ETA Model] --> Predict time for each road segment
#   |
# [Route-Level Aggregation] --> Sum segment ETAs + transition penalties
#   |
# [Post-Processing] --> Calibration, confidence intervals
#   |
# [ETA Response: 23 min (21-27 min range)]
#
# In-ride updates:
# [GPS Telemetry] --> [Remaining Route] --> [Updated ETA]

Step 3: Deep Dive — Spatial-Temporal Features

Road Segment Features (Static)

Feature	Type	Description
segment_length_m	Numerical	Length of road segment in meters
road_type	Categorical	Highway, arterial, residential, unpaved
speed_limit_kmh	Numerical	Posted speed limit
num_traffic_lights	Numerical	Number of traffic signals on segment
num_lanes	Numerical	Number of lanes in travel direction
has_turn	Binary	Whether segment involves a turn (adds delay)
elevation_change	Numerical	Elevation difference (uphill is slower)

Real-Time Features

Feature	Type	Description
current_speed_ratio	Numerical	Current avg speed / free-flow speed on segment
congestion_level	Categorical	Free flow, light, moderate, heavy, standstill
active_incidents	Binary	Accidents, road closures on or near segment
weather_condition	Categorical	Clear, rain, snow, fog (affects driving speed)
precipitation_mm	Numerical	Current rainfall/snowfall intensity
nearby_ride_speeds	Numerical	Average speed of rides currently on nearby segments

Temporal Features

Feature	Type	Description
hour_of_day	Cyclical (sin/cos)	Rush hour patterns: 7–9am, 5–7pm are slowest
day_of_week	Categorical	Weekday vs. weekend traffic patterns differ significantly
is_holiday	Binary	Holidays have different traffic patterns
minutes_since_event	Numerical	Post-event congestion near stadiums, concert venues
historical_segment_time	Numerical	Median travel time for this segment at this hour/day

Deep Dive — Model Architecture

Approach: Segment-Level Prediction + Aggregation

Rather than predicting trip ETA directly, predict the travel time for each road segment and sum them. This approach is more accurate because:

• Each segment has its own traffic pattern (highway vs. residential)
• The model generalizes to new routes it has never seen (novel combinations of known segments)
• You can update segment-level predictions in real-time as traffic changes

# Model architecture
#
# For each road segment in the route:
#   segment_features = [static_features, real_time_features, temporal_features]
#   segment_eta = model.predict(segment_features)  # seconds
#
# Route ETA = sum(segment_etas) + sum(transition_penalties)
#
# transition_penalty = f(turn_angle, traffic_light, stop_sign)
#
# Model choices:
# V1: Gradient Boosted Trees (XGBoost/LightGBM)
#     - Fast inference, handles mixed feature types
#     - Predict log(travel_time) to handle skewed distribution
#
# V2: Graph Neural Network (GNN)
#     - Model road network as a graph (intersections = nodes, roads = edges)
#     - Message passing captures traffic propagation effects
#     - A congested segment slows adjacent segments too
#     - Significantly more complex but captures spatial dependencies

Graph Neural Network Deep Dive

# GNN for road network ETA
#
# Graph construction:
#   Nodes = road intersections (~10M nodes for a city)
#   Edges = road segments connecting intersections
#   Node features = [traffic_signal, intersection_type, ...]
#   Edge features = [segment_length, speed_limit, current_speed, ...]
#
# Architecture:
#   1. Initial embedding: Linear(edge_features) --> h_0
#   2. Message passing (3 layers):
#      h_l+1 = GRU(h_l, aggregate(neighbor_messages))
#   3. Edge prediction: MLP(concat(h_source, h_target, edge_features))
#      --> predicted_travel_time for each segment
#
# Why GNN works:
#   - Captures traffic propagation (congestion on I-95 affects exits)
#   - Handles variable-length routes naturally
#   - Shared parameters across all segments (generalizes to new routes)

Deep Dive — Real-Time Updates

During a ride, the ETA must update as conditions change.

In-Ride ETA Update Flow

# Every 30 seconds during the ride:
#
# 1. Receive GPS telemetry from driver's phone
# 2. Map-match GPS to road segment (snap to nearest road)
# 3. Calculate actual speed on current segment
# 4. Compare actual progress vs. predicted progress
# 5. If behind schedule:
#    - Remaining segments get current traffic features refreshed
#    - Re-predict remaining segment ETAs
#    - New ETA = elapsed_time + sum(remaining_segment_etas)
# 6. If a reroute occurred:
#    - Compute ETA for new route from scratch
# 7. Smooth the ETA update (no sudden jumps):
#    new_displayed_eta = 0.7 * new_prediction + 0.3 * previous_eta

Metrics & Evaluation

Offline Metrics

Metric	Formula	Target
MAPE	mean(|actual - predicted| / actual)	< 12%
MAE	mean(|actual - predicted|) in seconds	< 120s for trips under 30 min
P80 Absolute Error	80th percentile of |actual - predicted|	< 120s
P95 Absolute Error	95th percentile of |actual - predicted|	< 300s
Under-prediction Rate	Fraction of trips where predicted < actual	< 40% (prefer over-predict)

Online Metrics

Metric	Description	Guardrail
Fare accuracy	Predicted fare vs. actual fare (ETA drives pricing)	Refund requests should not increase
User satisfaction	Post-ride ratings correlated with ETA accuracy	Should not decrease
Driver acceptance rate	Drivers accept rides based on estimated duration	Should not decrease
ETA convergence	How quickly in-ride ETA converges to actual	Within 10% by halfway point

Step 4: Trade-Offs & Extensions