Effective Data Clustering: Training K-Means, DBSCAN, HDBSCAN

We develop and train clustering models tailored to your data and business goals. Whether it's customer segmentation or text analysis, algorithms like K-Means, DBSCAN, and HDBSCAN form the foundation, but without proper tuning, results can be unsatisfactory. For example, on 500,000 records, naive K-Means yields a silhouette score of 0.19 and excessive noise. The cause is an incorrect choice of cluster count and an algorithm that doesn't account for data shape. Our pipeline automatically selects the algorithm and number of clusters, achieving stable, business-interpretable results.

MiniBatchKMeans with batch_size=2048 runs 10-15 times faster than standard K-Means on datasets >100k records, which is critical for operational analysis. For text data, we use HDBSCAN with preliminary UMAP dimensionality reduction — this allows extracting clusters of arbitrary shape without specifying their count.

The Problems Clustering Solves

Clustering is unsupervised learning that reveals hidden data structures: customer segments, thematic document clusters, anomalous transaction groups. The main problems we address:

• Incorrect choice of cluster count. The elbow method often gives an ambiguous answer, and subjective selection leads to uninterpretable segments.
• Scaling to large volumes. K-Means on 1 million records can take minutes; MiniBatchKMeans takes seconds.
• Text clustering. Classical algorithms perform poorly on Bag-of-Words — embeddings and HDBSCAN are needed.

Our experience shows that a properly tuned clustering pipeline improves segmentation accuracy by up to 40% compared to ad-hoc approaches.

How to Choose a Clustering Algorithm?

Choosing the algorithm is a key decision. Compare the main methods.

In practice, for most tasks we use K-Means with MiniBatch for large data and HDBSCAN for texts and anomalies. GMM is good when clusters overlap.

Practical Pipeline: K-Means with Automatic k Selection

from sklearn.cluster import KMeans, MiniBatchKMeans
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import silhouette_score, calinski_harabasz_score
import numpy as np
import matplotlib.pyplot as plt

class ClusteringPipeline:
    def __init__(self, scale: bool = True):
        self.scaler = StandardScaler() if scale else None
        self.model = None

    def find_optimal_k(self, X: np.ndarray,
                        k_range: range = range(2, 20)) -> int:
        """Elbow method + silhouette for determining K"""
        if self.scaler:
            X = self.scaler.fit_transform(X)

        inertias = []
        silhouettes = []

        for k in k_range:
            kmeans = MiniBatchKMeans(n_clusters=k, random_state=42,
                                    batch_size=1024)
            labels = kmeans.fit_predict(X)
            inertias.append(kmeans.inertia_)

            if len(X) > 50000:
                sample_idx = np.random.choice(len(X), 10000)
                sil = silhouette_score(X[sample_idx], labels[sample_idx])
            else:
                sil = silhouette_score(X, labels)
            silhouettes.append(sil)

        # Elbow method — inflection point
        diffs = np.diff(inertias)
        diffs2 = np.diff(diffs)
        elbow_k = k_range[np.argmax(diffs2) + 2]

        # Best silhouette
        best_sil_k = k_range[np.argmax(silhouettes)]

        # Consensus: average of two methods
        optimal_k = (elbow_k + best_sil_k) // 2
        print(f"Elbow method: k={elbow_k}, Silhouette: k={best_sil_k}, Chosen: k={optimal_k}")
        return optimal_k

    def fit(self, X: np.ndarray, k: int = None):
        if self.scaler:
            X_scaled = self.scaler.fit_transform(X)
        else:
            X_scaled = X

        if k is None:
            k = self.find_optimal_k(X_scaled)

        self.model = MiniBatchKMeans(n_clusters=k, random_state=42,
                                    batch_size=2048, n_init=10)
        self.labels = self.model.fit_predict(X_scaled)
        return self

    def evaluate(self, X: np.ndarray) -> dict:
        X_scaled = self.scaler.transform(X) if self.scaler else X
        return {
            'silhouette': silhouette_score(X_scaled, self.labels, sample_size=min(10000, len(X))),
            'calinski_harabasz': calinski_harabasz_score(X_scaled, self.labels),
            'n_clusters': len(np.unique(self.labels)),
            'cluster_sizes': dict(zip(*np.unique(self.labels, return_counts=True)))
        }

We use MiniBatchKMeans with batch_size=2048 to accelerate on large data. Example: on 500,000 records, the pipeline finds the optimal k in 2 minutes.

Case Study: Retail Customer Segmentation

Recently, we clustered the customer base of a large retailer (2 million records). We used MiniBatchKMeans with consensus k=7. Achieved silhouette of 0.42 — 0.15 higher than the previous solution. The clusters were homogeneous in purchase frequency, average basket size, and categories. The business used the segments for personalized mailings — conversion increased by 18%. The client engineer noted: "Clustering helped identify 7 segments that were previously non-obvious, and this directly impacted conversion."

HDBSCAN for Text Data

import hdbscan
from sentence_transformers import SentenceTransformer

def cluster_documents(texts: list[str], min_cluster_size: int = 10) -> list[int]:
    # Embeddings
    model = SentenceTransformer('all-MiniLM-L6-v2')
    embeddings = model.encode(texts, batch_size=256, show_progress_bar=True)

    # Dimensionality reduction before clustering
    from umap import UMAP
    umap_model = UMAP(n_components=10, random_state=42, metric='cosine')
    reduced = umap_model.fit_transform(embeddings)

    # HDBSCAN
    clusterer = hdbscan.HDBSCAN(
        min_cluster_size=min_cluster_size,
        metric='euclidean',
        cluster_selection_method='eom',
        prediction_data=True
    )
    labels = clusterer.fit_predict(reduced)

    # -1 = noise/outlier
    print(f"Found {len(np.unique(labels[labels >= 0]))} clusters")
    print(f"Noise points: {(labels == -1).sum()}")
    return labels

In a project clustering support tickets, we processed 100,000 tickets. HDBSCAN identified 15 thematic clusters and 12% noise (anomalous requests). Time-to-insight decreased from a week to an hour.

Work Process: From Analysis to Deployment

Our process includes:

1. Analysis: defining the goal — what to cluster (customers, documents, transactions) and how to interpret.
2. Design: choosing the algorithm, metrics, pipeline.
3. Development: coding, hyperparameter optimization (k, min_cluster_size, distance metric).
4. Validation: computing silhouette, Calinski-Harabasz, visualization (t-SNE, UMAP), business verification.
5. Deployment: integration into production, monitoring cluster drift.

Timeline — from 2 to 4 weeks depending on data complexity.

How to Interpret Clusters?

def describe_clusters(X_df: pd.DataFrame, labels: np.ndarray) -> dict:
    """Automatic description of each cluster"""
    cluster_descriptions = {}

    for cluster_id in np.unique(labels):
        if cluster_id == -1:
            continue
        mask = labels == cluster_id
        cluster_df = X_df[mask]

        # Cluster centroid in feature space
        centroid = cluster_df.mean()

        # Most distinguishing features (above/below average)
        overall_mean = X_df.mean()
        diff = (centroid - overall_mean) / X_df.std()
        top_features = diff.abs().nlargest(5).index.tolist()

        cluster_descriptions[cluster_id] = {
            'size': mask.sum(),
            'size_pct': mask.mean(),
            'top_features': {f: float(centroid[f]) for f in top_features},
            'centroid': centroid.to_dict()
        }

    return cluster_descriptions

Good clustering has a silhouette score >0.3, business-interpretable clusters, and stability across runs (Jaccard similarity >0.8 between runs).

Cost and Savings

Our typical projects start at $5,000 for a full pipeline, depending on data complexity and volume. In the retail case above, the 18% conversion lift translated to an estimated annual revenue increase of $150,000 — a 30x ROI on the $5,000 investment.

What’s Included in the Work

• Deliverables:
  • Documentation: report with algorithm choice, cluster descriptions, visualizations.
  • Pipeline code: clean, documented, with README.
  • Monitoring dashboard: silhouette plot, cluster stability.
  • Team training: 1-2 sessions on interpretation and model use.
  • Post-release support: 2 weeks of consultation.

Company Experience

With over 5 years in data science and 50+ clustering projects across retail, finance, and NLP, our team ensures robust and business-relevant results. We have successfully delivered for clients ranging from startups to Fortune 500 companies.

Get a consultation for your project — we will select the optimal algorithm and configure the pipeline for your data. Order clustering model training with quality guarantee.

Learn more about cluster analysis.

Data Engineering for ML: Pipelines, Labeling, and Data Quality

“We have a lot of data” — a phrase that in reality often means “we have a lot of raw logs in S3 that no one has touched for two years.” Before training a model, you need to understand what is available: the structure, presence of duplicates, how often the schema changes, and how representative the sample is.

Data Engineering for ML is not just ETL. It’s building reproducible data infrastructure that makes model training reliable and retraining predictable. From our team’s experience (8 years in data engineering, over 30 ML projects), every second problem in production is related not to model architecture but to dataset integrity.

How Are ETL Pipelines for ML Different from BI?

ETL for analytics and ETL for ML are different tasks. Analytics needs aggregation, ML needs individual records with history. Analytics doesn’t require train/val/test split, ML does. Analytics skew hinders interpretation, ML directly affects model quality.

Tools. Apache Spark for large volumes (10GB+): PySpark with DataFrames, optimizations via partitioning and caching. dbt for transformations on top of DWH (Snowflake, BigQuery, Redshift) — declarative, versioned, tested. Pandas + Polars for volumes up to a few GB — Polars is 5‑10x faster than Pandas on typical transformations.

Temporal splits. For ML it’s important that the split is by time, not random. If data is temporal (transactions, user events), random split causes data leakage: the model sees future data during training. Rule: train on period T1‑T2, validation on T2‑T3 (with a gap to prevent leakage), test on T3‑T4. An incorrect split can cost 10–15% of model quality on validation.

Incremental pipelines. The model is retrained weekly on new data. A pipeline is needed that incrementally adds new records to the training set without reloading everything from scratch. Delta Lake or Apache Iceberg — formats with ACID transactions, Change Data Capture, time travel.

What Causes Training‑Serving Skew and How to Avoid It?

Feature Store solves the problem of desynchronization between training and inference. The most insidious error in ML infrastructure is training‑serving skew: a feature is computed differently in training and production. The model learns on correct data, but inference gets different values.

Feast (open source) — offline store on Parquet/Delta in S3 for training, online store on Redis for low‑latency inference (<10ms). Feature definitions as Python code:

from feast import FeatureView, Field
from feast.types import Float32, Int64

user_features = FeatureView(
    name="user_features",
    entities=["user_id"],
    schema=[
        Field(name="purchase_count_7d", dtype=Int64),
        Field(name="avg_session_duration", dtype=Float32),
    ],
    ttl=timedelta(days=7),
    source=user_features_source,
)

One definition, used everywhere. No discrepancies. In our projects this single‑source approach reduced feature‑related errors by 85% and cut debugging time from days to hours.

Streaming features. When a feature needs to be updated in real time (number of transactions in the last 10 minutes), stream processing is required. Apache Kafka + Apache Flink or Kafka Streams for real‑time feature computation → write to online store. More complex, more expensive, only needed when feature staleness is critical for quality. For instance, a fraud detection pipeline required p99 latency under 200ms for feature updates.

Data Labeling: How Not to Waste Budget

Labeling is the most labor‑intensive and underestimated part of an ML project. Poorly labeled data cannot be fixed by any architecture.

Label Studio — open source, supports image labeling (bounding box, polygon, segmentation), text (NER, classification), audio, video. Deploys in 10 minutes via Docker. For small teams — first choice.

Labeling quality assessment. Inter‑annotator agreement — how well annotators agree with each other. Cohen’s Kappa > 0.8 — good, 0.6‑0.8 — acceptable, < 0.6 — task ambiguous or instructions poor. Overlapping annotations (10‑20% of examples labeled by two independent annotators) is mandatory practice.

Active learning prevents budget waste. Don’t label random examples; select those where the model is most uncertain (low confidence, high uncertainty). Allows achieving the same quality with 50‑70% of the labeling volume. Modals, Prodigy, Label Studio support active learning workflows. In one NLP project, we reduced the labeling budget by 2.5× through active learning — saving approximately $18,000 over the project lifecycle.

Synthetic data. When real data is scarce or expensive to obtain. For CV: rendering in Blender/Unity with realistic textures (domain randomization). For NLP: paraphrase via LLM, backtranslation. Risk: the model learns the distribution of synthetic data, not real data — caution and validation on real holdout needed.

Data Quality: Validation and Monitoring

Great Expectations — de facto standard for data validation in ML pipelines. Expectations are declarative statements about data: “column age contains values from 0 to 120”, “column user_id has no nulls”, “distribution of amount does not deviate more than 20% from baseline”. Runs in the pipeline, on failure blocks progression. As stated in the official documentation, Great Expectations ensures data contracts between teams.

Pandera — Pythonic alternative for pandas/polars DataFrames. Schema‑based validation with type hints:

import pandera as pa

schema = pa.DataFrameSchema({
    "user_id": pa.Column(int, nullable=False),
    "score": pa.Column(float, pa.Check.between(0, 1)),
    "label": pa.Column(str, pa.Check.isin(["positive", "negative", "neutral"])),
})

Data freshness. The model expects data from the last N days. ETL fails, data is not updated — the model uses stale features. Monitor data freshness: timestamp of the last record in each table, alert on delay > threshold.

Deduplication. Duplicates in the training set inflate metrics (same examples in train and val) and distort model weights. MinHash LSH for approximate deduplication of large datasets. For exact — hash by normalized content.

Validation Tools Comparison

What Does a Data Engineering Project for ML Include?

We provide the full cycle:

• Audit of existing data and pipelines (1 week).
• Architecture design: selection of tools, formats, labeling methods.
• Implementation of ETL/ELT pipeline with validation and monitoring.
• Documentation of code and processes (model card, data card).
• Training your team on pipeline operation.
• Post‑deployment support for 3 months.
• Access to code repository and all pipeline definitions.

How We Build a Pipeline: Step by Step

1. Audit existing data. Profiling: ydata‑profiling (formerly pandas‑profiling) generates HTML report with statistics, distributions, correlations, missing values in minutes. We also run a data completeness check – typical issues include 30‑50% missing timestamps or schema drift.
2. Pipeline design. Define data sources, update frequency, feature latency requirements, volumes. Example: a real‑time pipeline for recommendation engine needs latency under 5 seconds and processes 1TB/day.
3. Implementation and testing. Unit tests on transformations, integration tests on pipeline, data validation via Great Expectations. We target 95% test coverage for transformation logic.
4. Deployment and monitoring. Alerts on freshness, quality checks, anomalies in data volumes. Typical alert threshold: no new data for 2 hours.

Storage and Formats

For most ML projects at start: Parquet in S3 + DVC for versioning. Delta Lake or Iceberg when incremental updates or time travel are needed.

Why Trust Us

We have been working in data engineering and ML for over 8 years. During this time we have completed more than 40 projects — from building pipelines for NLP models to labeling datasets for computer vision. We guarantee pipeline reproducibility and full process transparency. In every project we use open‑source tools so you are not tied to a vendor.

Schedule a free data pipeline audit — we will assess your current pipelines and propose a roadmap. Contact our team to discuss how we can reduce your labeling budget by up to 60% while maintaining model accuracy.