Memes look chaotic. They're not.

Beneath the humor, irony, and internet slang lies a surprisingly consistent taxonomy - patterns of genre, tone, cultural reference, and emotional register that repeat across thousands of posts. This project sets out to reverse-engineer that hidden structure using a fully interpretable NLP pipeline, built from scratch with standard Python libraries.

"Meme culture is not noise - it's compressed cultural signal. Our job is to decompress it."

The goal: cluster 5,819 meme descriptions into meaningful, human-interpretable thematic groups - using no external datasets, no pre-trained models, and no AutoML.

The competition dataset comprises 5,819 meme records, each described by three fields.

The primary analytical signal is the input column. Each entry follows an implicit structure of the form: caption → scene description → poster intent → entity mapping providing multi-layered semantic information within a single text field. The url column was used supplementarily to extract domain-level metadata that could hint at community context.

The end-to-end pipeline follows a sequential transformation from raw text to labelled clusters:

Raw text descriptions were cleaned through the following sequential steps:

Two complementary representations were constructed from the cleaned text.

TF-IDF Vectorization

Bigrams preserve meaningful multi-word concepts like "meme poster", "trying convey", and "not related" that carry structural meaning a unigram model would miss.

Latent Semantic Analysis (LSA)

LSA projects the 2,000-dimensional sparse TF-IDF space into a dense 50-dimensional semantic space, capturing latent co-occurrence patterns that pure term matching misses entirely.

The optimal number of clusters was determined empirically using two independent metrics evaluated across K = 4 to K = 14:

• Elbow Method → Inertia flattens at the inflection point
• Silhouette Score → Peaks at the optimal cluster count

→ K = 7 selected, yielding clusters large enough for statistical stability and specific enough for thematic coherence.

Primary Method → K-Means

K-Means clustering was applied to the L2-normalized LSA vectors with K = 7, using 20 random initializations to reduce sensitivity to centroid initialization.

Validation Method → Agglomerative Clustering (Ward Linkage)

Using two fundamentally different algorithms (centroid-based vs hierarchical) and measuring their agreement via Adjusted Rand Index (ARI) provides a strong statistical argument that the discovered clusters are real structure, not noise.

t-SNE was applied to the 50-dimensional LSA vectors to produce a 2-dimensional projection for visual inspection. The resulting scatter plot, saved as cluster_viz.png, confirms meaningful spatial separation between cluster regions, corroborating the quantitative validation results.

Cluster labels were assigned through a two-step interpretive process:

1. Keyword extraction - A per-cluster TF-IDF analysis was run to identify the top 10 most distinctive terms for each group
2. Manual review - A minimum of 15 meme entries per cluster were manually read to validate that keyword-derived labels accurately reflected the full range of content within each group

Seven thematic clusters were identified, each with a human-assigned label grounded in keyword analysis and manual content review.

All cluster labels emerged organically from keyword convergence and content inspection (no labels were pre-defined)

Every component of this pipeline is fully reproducible:

• All random seeds fixed to 42
• Notebook executes top-to-bottom without errors
• Model artifacts saved as .pkl for independent inspection
• No external data downloads or API calls required
• output.txt generated programmatically from the notebook

The following improvements are proposed conceptually. Here's what an unconstrained v2.0 could look like:

The url column points to actual meme images. A future system could combine OCR (to extract overlaid text) with image classifiers (to identify visual meme templates like the Drake format or Distracted Boyfriend). Visual + textual fusion would dramatically improve cluster fidelity.

An emotional layer (positive / negative / sarcastic / dark) would split clusters that share a theme but differ in tone — separating "motivational memes" from "motivational sarcasm", for example.

A two-level system would mirror how humans naturally categorize memes:

Humor    →  [Dark Humor, Self-deprecating, Absurdist]
Culture  →  [Gaming, Anime, Sports, Politics]
Emotion  →  [Wholesome, Relatable, Rage-bait]

With timestamps, tracking cluster growth and decay over time could reveal meme lifecycle patterns and predict emerging formats before they go viral.

URL domains (Reddit subreddits, Know Your Meme, Imgur) encode deep community context. Graph-based clustering that weights memes from the same origin community could significantly improve thematic coherence.

A semi-supervised approach where domain experts label a seed set of 50–100 memes could bootstrap constrained K-Means or label propagation — combining machine scale with human judgment.

• Content Moderation — Automatically flag offensive meme clusters for human review
• Recommendation Engines — Suggest memes from contextually related clusters
• Cultural Analytics — Track the evolution of internet humor across communities and time

Meme Context Clustering
├── NOTICE
├── README.md
├── LICENSE
├── data/ 
    └── meme.csv           -> Original Data
├── models/ 
    ├── vectorizer.pkl     -> Trained TF-IDF vectorizer
    └── cluster_model.pkl  -> Trained K-Means model
├── report.txt             -> Whole Approach report
├── output.txt             -> Cluster listings + per-meme reasoning + methodology + pitch
└── meme_clustering.ipynb  -> Full reproducible pipeline notebook