Duration Prediction
Business Case
Problem Statement
Project duration estimation challenges:
- Subjective expert estimates
- Lack of historical benchmarking
- Inaccurate early-stage predictions
- Difficulty comparing similar projects
Solution
Machine learning-based duration prediction using k-Nearest Neighbors and regression models trained on historical project data.
Technical Implementation
import pandas as pd
import numpy as np
from typing import Dict, Any, List, Optional, Tuple
from dataclasses import dataclass, field
from datetime import date
from enum import Enum
import math
class ModelType(Enum):
KNN = "knn"
LINEAR_REGRESSION = "linear_regression"
WEIGHTED_KNN = "weighted_knn"
class ProjectType(Enum):
OFFICE = "office"
RESIDENTIAL = "residential"
INDUSTRIAL = "industrial"
RETAIL = "retail"
HEALTHCARE = "healthcare"
EDUCATION = "education"
@dataclass
class ProjectFeatures:
project_id: str
project_type: ProjectType
size_sf: float
floors: int
complexity: int # 1-5
location_factor: float # Cost adjustment factor
has_basement: bool = False
is_renovation: bool = False
actual_duration: Optional[int] = None # Days
@dataclass
class PredictionResult:
predicted_duration: int
confidence_interval: Tuple[int, int]
similar_projects: List[str]
model_used: ModelType
features_importance: Dict[str, float]
class DurationPredictor:
"""Predict project duration using ML techniques."""
def __init__(self):
self.training_data: List[ProjectFeatures] = []
self.feature_weights: Dict[str, float] = {
'size_sf': 0.30,
'floors': 0.15,
'complexity': 0.25,
'location_factor': 0.10,
'has_basement': 0.10,
'is_renovation': 0.10
}
self.type_baseline_days: Dict[ProjectType, Dict[str, float]] = {
ProjectType.OFFICE: {'base': 300, 'per_1000sf': 0.5},
ProjectType.RESIDENTIAL: {'base': 240, 'per_1000sf': 0.4},
ProjectType.INDUSTRIAL: {'base': 180, 'per_1000sf': 0.3},
ProjectType.RETAIL: {'base': 200, 'per_1000sf': 0.35},
ProjectType.HEALTHCARE: {'base': 400, 'per_1000sf': 0.6},
ProjectType.EDUCATION: {'base': 320, 'per_1000sf': 0.45}
}
def add_training_project(self, project: ProjectFeatures):
"""Add project to training dataset."""
if project.actual_duration is not None:
self.training_data.append(project)
def load_training_data(self, df: pd.DataFrame):
"""Load training data from DataFrame."""
for _, row in df.iterrows():
project = ProjectFeatures(
project_id=str(row['project_id']),
project_type=ProjectType(row['project_type'].lower()),
size_sf=float(row['size_sf']),
floors=int(row['floors']),
complexity=int(row['complexity']),
location_factor=float(row.get('location_factor', 1.0)),
has_basement=bool(row.get('has_basement', False)),
is_renovation=bool(row.get('is_renovation', False)),
actual_duration=int(row['actual_duration'])
)
self.add_training_project(project)
def _extract_features(self, project: ProjectFeatures) -> np.ndarray:
"""Extract feature vector from project."""
return np.array([
project.size_sf / 10000, # Normalize to 10k SF
project.floors,
project.complexity,
project.location_factor,
1 if project.has_basement else 0,
1 if project.is_renovation else 0
])
def _calculate_distance(self, features1: np.ndarray,
features2: np.ndarray) -> float:
"""Calculate weighted Euclidean distance."""
weights = np.array(list(self.feature_weights.values()))
diff = (features1 - features2) ** 2
weighted_diff = diff * weights
return math.sqrt(np.sum(weighted_diff))
def _find_k_nearest(self, target: ProjectFeatures, k: int = 5,
same_type: bool = True) -> List[Tuple[ProjectFeatures, float]]:
"""Find k nearest neighbors."""
target_features = self._extract_features(target)
distances = []
for project in self.training_data:
if same_type and project.project_type != target.project_type:
continue
proj_features = self._extract_features(project)
distance = self._calculate_distance(target_features, proj_features)
distances.append((project, distance))
distances.sort(key=lambda x: x[1])
return distances[:k]
def predict_knn(self, target: ProjectFeatures, k: int = 5) -> PredictionResult:
"""Predict duration using k-NN."""
nearest = self._find_k_nearest(target, k)
if not nearest:
# Fall back to baseline
return self._predict_baseline(target)
# Simple average of k nearest
durations = [p.actual_duration for p, _ in nearest]
predicted = int(np.mean(durations))
# Confidence interval (using std dev)
std = np.std(durations)
lower = int(predicted - 1.96 * std)
upper = int(predicted + 1.96 * std)
return PredictionResult(
predicted_duration=predicted,
confidence_interval=(max(1, lower), upper),
similar_projects=[p.project_id for p, _ in nearest],
model_used=ModelType.KNN,
features_importance=self.feature_weights
)
def predict_weighted_knn(self, target: ProjectFeatures, k: int = 5) -> PredictionResult:
"""Predict duration using distance-weighted k-NN."""
nearest = self._find_k_nearest(target, k)
if not nearest:
return self._predict_baseline(target)
# Inverse distance weighting
total_weight = 0
weighted_sum = 0
for project, distance in nearest:
weight = 1 / (distance + 0.001) # Add small value to avoid division by zero
weighted_sum += project.actual_duration * weight
total_weight += weight
predicted = int(weighted_sum / total_weight)
# Confidence interval
durations = [p.actual_duration for p, _ in nearest]
std = np.std(durations)
lower = int(predicted - 1.96 * std)
upper = int(predicted + 1.96 * std)
return PredictionResult(
predicted_duration=predicted,
confidence_interval=(max(1, lower), upper),
similar_projects=[p.project_id for p, _ in nearest],
model_used=ModelType.WEIGHTED_KNN,
features_importance=self.feature_weights
)
def predict_regression(self, target: ProjectFeatures) -> PredictionResult:
"""Predict duration using linear regression."""
if len(self.training_data) < 3:
return self._predict_baseline(target)
# Filter by project type
same_type = [p for p in self.training_data if p.project_type == target.project_type]
if len(same_type) < 3:
same_type = self.training_data
# Build feature matrix and target vector
X = np.array([self._extract_features(p) for p in same_type])
y = np.array([p.actual_duration for p in same_type])
# Simple linear regression using normal equations
X_with_intercept = np.column_stack([np.ones(len(X)), X])
try:
# beta = (X'X)^-1 X'y
XtX = X_with_intercept.T @ X_with_intercept
XtX_inv = np.linalg.inv(XtX)
beta = XtX_inv @ X_with_intercept.T @ y
except np.linalg.LinAlgError:
return self._predict_baseline(target)
# Predict
target_features = self._extract_features(target)
target_with_intercept = np.arr