AI-система матчинга инфлюенсеров и аналитики аудитории
Ручной поиск инфлюенсеров — дорого и ненадёжно. AI-матчинг анализирует не только тематику блогера, но и качество аудитории (боты, fake engagement), пересечение с целевой аудиторией бренда и прогнозирует ROI кампании. Платформы типа GRIN, Traackr, Upfluence используют именно эти подходы.
Аналитика качества аудитории
import numpy as np
import pandas as pd
from sklearn.ensemble import IsolationForest
from sklearn.cluster import KMeans
import json
from anthropic import Anthropic
class InfluencerAudienceAnalyzer:
"""Анализ качества и состава аудитории инфлюенсера"""
def compute_authenticity_score(self, account_data: dict) -> dict:
"""
Скор аутентичности аудитории (0-100).
Детектирование ботов и искусственного engagement.
"""
followers = account_data.get('followers_count', 1)
avg_likes = account_data.get('avg_likes', 0)
avg_comments = account_data.get('avg_comments', 0)
avg_views = account_data.get('avg_views', followers)
# Engagement Rate (ER)
er = (avg_likes + avg_comments) / followers * 100
# Follower-to-Following ratio (аномалии = много ботов-подписчиков)
follow_ratio = account_data.get('followers_count', 1) / max(
account_data.get('following_count', 1), 1
)
# Рост аудитории (резкие скачки = накрутка)
growth_spike = account_data.get('max_weekly_growth_pct', 0)
# Views/Follower ratio для видео
views_ratio = avg_views / followers if followers > 0 else 0
score = 100.0
issues = []
# Слишком низкий ER (нормы: nano 5-10%, micro 3-6%, macro 1-3%, mega 0.5-1.5%)
size_tier = self._get_tier(followers)
expected_er_range = {'nano': (5, 10), 'micro': (3, 6), 'macro': (1, 3), 'mega': (0.5, 1.5)}
expected_range = expected_er_range.get(size_tier, (1, 5))
if er < expected_range[0] * 0.5:
score -= 30
issues.append(f'ER {er:.1f}% значительно ниже нормы {expected_range[0]}% для {size_tier}')
elif er < expected_range[0]:
score -= 15
# Аномально высокий ER (накрутка лайков)
if er > expected_range[1] * 3:
score -= 20
issues.append('Аномально высокий ER — возможна накрутка')
# Резкий рост
if growth_spike > 50:
score -= 25
issues.append(f'Резкий рост аудитории +{growth_spike:.0f}% за неделю')
# Низкое соотношение просмотров
if views_ratio < 0.1 and account_data.get('content_type') == 'video':
score -= 15
issues.append('Низкий охват видео-контента')
return {
'authenticity_score': max(0, round(score)),
'engagement_rate': round(er, 2),
'tier': size_tier,
'issues': issues,
'estimated_real_followers': int(followers * max(0, score) / 100)
}
def _get_tier(self, followers: int) -> str:
if followers < 10000:
return 'nano'
elif followers < 100000:
return 'micro'
elif followers < 1000000:
return 'macro'
return 'mega'
def analyze_audience_demographics(self, follower_sample: pd.DataFrame,
brand_target_audience: dict) -> dict:
"""Пересечение аудитории инфлюенсера с ЦА бренда"""
overlaps = {}
# Гендер
if 'gender' in follower_sample.columns and 'gender' in brand_target_audience:
brand_gender = brand_target_audience['gender']
influencer_gender_dist = follower_sample['gender'].value_counts(normalize=True).to_dict()
overlaps['gender_match'] = influencer_gender_dist.get(brand_gender, 0)
# Возраст
if 'age_group' in follower_sample.columns and 'age_groups' in brand_target_audience:
target_ages = set(brand_target_audience['age_groups'])
influencer_ages = set(
follower_sample['age_group'].value_counts(normalize=True)
.nlargest(3).index.tolist()
)
overlaps['age_overlap'] = len(target_ages & influencer_ages) / max(len(target_ages), 1)
# Геолокация
if 'country' in follower_sample.columns and 'countries' in brand_target_audience:
target_countries = set(brand_target_audience['countries'])
influencer_countries = set(
follower_sample['country'].value_counts(normalize=True)
.nlargest(5).index.tolist()
)
overlaps['geo_overlap'] = len(target_countries & influencer_countries) / max(len(target_countries), 1)
# Общий скор аффинности
overlaps['audience_affinity'] = round(np.mean(list(overlaps.values())) if overlaps else 0.5, 2)
return overlaps
class InfluencerMatcher:
"""Матчинг инфлюенсеров под кампанию бренда"""
def __init__(self):
self.llm = Anthropic()
self.analyzer = InfluencerAudienceAnalyzer()
def score_influencer(self, influencer: dict,
campaign: dict,
follower_sample: pd.DataFrame) -> dict:
"""Комплексный скор инфлюенсера для кампании"""
# Качество аудитории
authenticity = self.analyzer.compute_authenticity_score(influencer)
# Пересечение с ЦА
audience_match = self.analyzer.analyze_audience_demographics(
follower_sample, campaign.get('target_audience', {})
)
# Тематическое соответствие (категории контента)
content_categories = set(influencer.get('content_categories', []))
brand_categories = set(campaign.get('relevant_categories', []))
category_match = len(content_categories & brand_categories) / max(len(brand_categories), 1)
# Прогноз CPE (Cost Per Engagement)
budget_per_influencer = campaign.get('budget', 10000)
expected_engagements = (
influencer.get('followers_count', 0) *
authenticity['engagement_rate'] / 100 *
authenticity['authenticity_score'] / 100
)
cpe = budget_per_influencer / max(expected_engagements, 1)
# Итоговый скор
total_score = (
authenticity['authenticity_score'] / 100 * 0.30 +
audience_match.get('audience_affinity', 0.5) * 0.35 +
category_match * 0.25 +
min(1.0, 10 / max(cpe, 0.1)) * 0.10 # Инвертируем CPE (меньше = лучше)
)
return {
'influencer_id': influencer.get('id'),
'handle': influencer.get('handle'),
'tier': authenticity['tier'],
'total_score': round(total_score, 3),
'authenticity': authenticity['authenticity_score'],
'audience_affinity': audience_match.get('audience_affinity', 0),
'category_match': round(category_match, 2),
'expected_engagements': int(expected_engagements),
'estimated_cpe': round(cpe, 2),
'red_flags': authenticity['issues']
}
def generate_campaign_brief(self, influencer: dict,
campaign: dict) -> str:
"""Персональный бриф для инфлюенсера"""
response = self.llm.messages.create(
model="claude-3-5-sonnet-20241022",
max_tokens=300,
messages=[{
"role": "user",
"content": f"""Write a personalized campaign brief for an influencer in Russian.
Influencer: @{influencer.get('handle')}, {influencer.get('tier')} tier, {influencer.get('content_categories', [])} content
Campaign: {campaign.get('name')}, brand: {campaign.get('brand_name')}
Product: {campaign.get('product_description', '')}
Key message: {campaign.get('key_message', '')}
Target audience: {campaign.get('target_audience', {})}
Write a 2-3 paragraph brief that:
1. Explains why this specific influencer was chosen (personalized)
2. Describes the campaign goals and what we want to achieve
3. Gives creative guidelines that fit their style"""
}]
)
return response.content[0].text
AI-матчинг инфлюенсеров снижает CPE на 25-40% по сравнению с ручным отбором за счёт точного аудиторного пересечения. Главный ROI-драйвер — exclusion ботов: 30-60% аудитории типичного macro-инфлюенсера могут составлять неактивные или фейковые аккаунты.