FashionTech AI
Fashion is one of the most volatile industries: trends change within weeks, excess inventory goes on sale at 60-70% off, and stock-outs of essential items cost lost sales. AI is redesigning processes from trend forecasting to personalized shopping.
Trend forecasting
Trend Intelligence from Social Media:
Runways and the street. Instagram, TikTok, Pinterest—a real fashion barometer:
import requests
import pandas as pd
from transformers import pipeline
from sklearn.decomposition import LatentDirichletAllocation
from sklearn.feature_extraction.text import CountVectorizer
import numpy as np
class FashionTrendAnalyzer:
"""Анализ трендов из социальных медиа для модной индустрии"""
def __init__(self):
self.image_classifier = pipeline(
'image-classification',
model='patrickjohncyh/fashion-clip' # CLIP fine-tuned on fashion
)
self.trend_categories = [
'oversized_silhouette', 'minimalism', 'bold_colors',
'pattern_mixing', 'monochromatic', 'vintage', 'streetwear',
'sustainable_fabrics', 'gender_neutral'
]
def classify_fashion_image(self, image):
"""Классификация модного образа по трендам"""
results = self.image_classifier(image,
candidate_labels=self.trend_categories)
return {r['label']: r['score'] for r in results}
def track_trend_velocity(self, trend_scores_history):
"""
Скорость роста/падения тренда.
trend_scores_history: DataFrame [date × trend] с агрегированными score
"""
velocities = {}
for trend in self.trend_categories:
if trend in trend_scores_history.columns:
series = trend_scores_history[trend]
# Линейный тренд за последние 4 недели
x = np.arange(len(series))
slope = np.polyfit(x, series, 1)[0]
velocities[trend] = {
'current_score': float(series.iloc[-1]),
'weekly_change': float(slope * 7),
'direction': 'rising' if slope > 0 else 'falling',
'weeks_to_peak': max(0, (1.0 - series.iloc[-1]) / slope) if slope > 0 else 0
}
return velocities
Lifecycle trend forecast:
Each trend goes through: Emerging → Growing → Peak → Declining. The ML model determines the phase based on the growth rate and saturation: - Emerging: sharp growth from scratch → early buyers - Peak: slowing growth → mass market - Declining → stop orders
Demand forecasting
Attribute-based Forecasting:
Forecast not by SKU (too short history), but by attributes: - Features: color, silhouette, material, trend-belonging, price segment - Model: hierarchical forecast (category → subcategory → attribute → SKU) - Cold start: new article → forecast by similar historical
from lightgbm import LGBMRegressor
def build_fashion_demand_model(sales_df, product_attributes):
"""
Прогноз продаж для SKU по атрибутам продукта.
Решает проблему cold start для новых коллекций.
"""
# Объединить продажи с атрибутами
df = sales_df.merge(product_attributes, on='sku_id')
feature_cols = [
# Атрибуты продукта
'color_group', 'silhouette', 'material', 'price_segment',
'trend_score', 'season',
# Временные признаки
'week_of_year', 'days_since_launch',
'promo_flag', 'new_arrival',
# История похожих SKU (same attributes)
'similar_sku_avg_sales_w1', 'similar_sku_avg_sales_w2',
]
model = LGBMRegressor(n_estimators=300, num_leaves=64)
model.fit(df[feature_cols], df['weekly_units'])
return model
Visual Search and Personalization
Visual Fashion Search:
"Find Similar" by Photo - A Killer Feature for Fashion: - Photo Upload → CLIP Embedding → Cosine Similarity in Product Catalog - Text Filter Supplement: "Similar, but Blue and Under 5,000 Rubles" - Retrieval Augmented: First, visually similar items, then ranking based on personal preferences
Outfit Completion:
What to choose for the selected item: - Graph neural network: products as nodes, outfit compatibility as edges - Training on a dataset of “successful” looks (high engagement outfits) - Constraints: price range, user style, availability in stock
Optimization of size chart and returns
Fit Prediction:
The most common reason for returns in fashion is size mismatch: - User details: height, weight, previous size returns - Brand details: fit model, size chart, "runs big/runs small" reviews - ML recommendation: "For your body type, we recommend XL, this brand runs small"
Result: a reduction in the return rate from 20–30% to 12–15% for online clothing sales.
Development time: 5–8 months for the FashionTech AI platform with trend intelligence, demand forecasting, visual search, and fit prediction.