Apply for AI Grants India

Why Word2Vec needs a different football dataset

Word2Vec is often introduced as a tool for finding similar words, but its more useful role here is to find similar player profiles, match situations, and tactical patterns. The model learns from context: terms appearing near one another repeatedly receive similar vector representations. In football, that context can come from scouting reports, match commentary, event descriptions, coach notes, or structured sequences converted into text.

That distinction matters in India. The Indian football ecosystem spans the Indian Super League, I-League, state competitions, youth academies, university football, and unevenly documented grassroots programmes. A model trained only on highly covered professional matches will favour players and clubs with better data. Treat Word2Vec as a discovery and comparison layer—not as a standalone rating system.

Teams building broader analytics pipelines can pair this workflow with Indian open-source AI developer projects to reduce infrastructure costs and adapt models locally.

Define the scouting question first

Start with a decision the model must support. Useful questions include:

• Which young full-backs resemble the team’s preferred overlapping profile?
• Which midfielders have comparable progression, pressing, and ball-retention patterns?
• Which players could transition from I-League or state-level football into a higher tempo competition?
• Which tactical roles are underrepresented in the current squad?

Avoid asking Word2Vec to identify the “best” player. Similarity is not quality. A low-cost defensive midfielder may resemble an established player in role and actions while differing substantially in consistency, opposition quality, fitness, or decision-making.

Build player documents from structured data

Word2Vec requires sequences of tokens. A useful approach is to create one or more documents per player, rather than placing all raw statistics into a spreadsheet and expecting the model to interpret them.

For every player, combine several evidence types:

• Role and context: position, preferred foot, age band, league, team, minutes, and competition level.
• Actions: progressive passes, carries, pressures, interceptions, aerial duels, chances created, shots, and defensive recoveries.
• Situations: transition, set piece, high press, low block, final-third possession, and defensive restarts.
• Scouting language: “underlapping full-back”, “press-resistant midfielder”, or “target forward”.
• Time: season, match phase, and recent form window.

Convert meaningful observations into tokens such as progressive_pass, right_half_space, counterpress, and aerial_duel. Do not create thousands of noisy tokens from every tiny statistical fluctuation. Bin metrics into interpretable bands—for example, high_progressive_pass_rate—and retain the underlying numbers for later validation.

A player document might look like this:

player_rahul_midfielder league_isl role_8
progressive_pass_high carry_into_final_third medium_press_resistance_high
counterpress frequent recovery_in_middle_third chance_creation_medium

player_arjun_midfielder league_i_league role_8
progressive_pass_high carry_into_final_third medium_press_resistance_high
counterpress frequent recovery_in_middle_third chance_creation_medium

Keep identity tokens such as player_rahul separate from descriptive tokens. Otherwise, a model can learn that two names appear in the same report without learning why their profiles are alike.

Prepare and split the corpus carefully

Clean spelling, punctuation, club abbreviations, and positional labels before training. Standardise variants such as centre-back, CB, and central_defender, but preserve distinctions that matter tactically. Hindi, Bengali, Malayalam, and other local-language scouting notes may contain valuable information; translate only when necessary and retain the original text for auditability.

Create a time-based validation split. Train on earlier seasons and test on a later season where possible. A random split can leak repeated reports, player names, or match contexts into both sets and make the model appear more accurate than it is. Also track coverage by league, gender, age group, and region so the output is not mistaken for a complete map of Indian talent.

Train a baseline Word2Vec model

Gensim is sufficient for a strong first experiment:

from gensim.models import Word2Vec

sentences = [
    ["role_8", "progressive_pass_high", "press_resistance_high", "counterpress"],
    ["role_8", "progressive_pass_high", "carry_final_third", "counterpress"],
    ["role_9", "box_occupancy_high", "aerial_duel_high", "shot_volume_medium"],
]

model = Word2Vec(
    sentences=sentences,
    vector_size=100,
    window=4,
    min_count=3,
    workers=4,
    sg=1,
    negative=10,
    epochs=30,
    seed=42,
)

print(model.wv.most_similar("role_8", topn=10))

Use Skip-gram (`sg=1`) when the corpus is modest and you care about less frequent tactical patterns. Use CBOW when you have a larger, repetitive corpus and need faster training. Test vector sizes such as 50, 100, and 200 rather than assuming a larger embedding is better. Set a seed, record parameters, and save the training corpus version.

For player comparison, aggregate the vectors of a player’s descriptive tokens or train player IDs in the same context as action tokens. Cosine similarity is a starting point, not a final verdict:

from numpy import mean
from sklearn.metrics.pairwise import cosine_similarity

profile = ["progressive_pass_high", "press_resistance_high", "counterpress"]
player_vector = mean([model.wv[token] for token in profile], axis=0)

In production, create player embeddings from multiple matches and weight observations by minutes, competition level, and recency. Never allow a two-match sample to carry the same confidence as a full season.

Turn embeddings into scouting workflows

Use the model to generate a shortlist, then verify it with football and operational constraints. A practical workflow is:

1. Query similar role or action tokens.
2. Retrieve players whose documents contain those patterns.
3. Filter by age, position, registration rules, salary range, location, and availability.
4. Compare raw metrics against players in the same league and role.
5. Review video and obtain coach or scout feedback.
6. Track whether the recommendation remains valid in a later sample.

Word2Vec can support role-based recruitment, succession planning, academy progression, and opponent analysis. It can also reveal that a player’s statistical profile is similar to a different position—for example, a wide midfielder whose actions resemble an attacking full-back. That is a prompt for investigation, not an automatic positional conversion.

Founders building scouting or recruitment products should also consider how their outputs fit into broader cost-effective recruitment platforms for Indian founders, especially when clubs need explainable shortlists rather than opaque scores.

Validate similarity instead of trusting the nearest neighbours

A useful evaluation set should include pairs labelled by experienced scouts or coaches: similar role, similar style, same position but different style, and clearly dissimilar players. Measure precision among the top five or ten recommendations, agreement with expert labels, and stability across model seeds and seasons.

Run ablation tests: remove names, club tokens, or league tokens and see whether recommendations still make tactical sense. If every player from one club clusters together, the model may be learning reporting style or team system rather than individual ability. Compare Word2Vec with simpler baselines such as z-score-normalised statistics, cosine similarity on event features, or clustering with PCA and k-means. A more complex method should earn its place through better decisions.

Common risks in Indian football data

• Uneven coverage: prominent clubs and men’s professional competitions may dominate the corpus.
• Small samples: youth and state-level players can appear in too few matches for reliable embeddings.
• Role ambiguity: a “midfielder” label may cover a No. 6, No. 8, and No. 10.
• Team effects: possession, pressing, and chance creation are shaped by teammates and coaching systems.
• Language and transcription noise: translated or auto-generated reports can alter tactical meaning.
• Privacy and consent: do not ingest private medical, disciplinary, or social data without a lawful basis and clear governance.

Keep an audit trail showing the data used, tokens contributing to a recommendation, confidence limits, and human decisions. Do not infer personality, injury risk, or employability from embeddings. Protect minors especially carefully, and ensure that regional, linguistic, and socioeconomic gaps do not become hidden selection criteria.

A practical 30-day pilot

In week one, define one recruitment question and assemble a small, documented corpus. In week two, standardise labels, create tokens, and train Word2Vec alongside a statistical baseline. In week three, have scouts review nearest-neighbour results blind to model scores. In week four, revise the corpus, test on a later time window, and publish a short model card covering coverage, limitations, and acceptable use.

The strongest implementation is usually modest: clean data, clear role definitions, time-aware validation, and a scout who can challenge every recommendation. Word2Vec can make Indian football data easier to search and compare, but the value comes from connecting embeddings to accountable recruitment and player-development decisions.