CS2 Player Role Classifier

BTP Mid-Term — Sachin Kumar | UG-2 | S20240010206

Measuring Player Role Consistency and Its Impact on Match Outcome in CS2 — using behavioural signals extracted from raw .dem replay files, without manual annotation.

---

What This Does

CS2 professional players fall into recognisable tactical roles — Entry fraggers who push first, Lurkers who flank late, Supports who flash and trade — but no demo file explicitly labels them. This pipeline extracts 14 per-round behavioural features from raw binary .dem recordings and uses an ensemble of three unsupervised clustering algorithms to assign each player a role automatically.

Two phases:

1. G1 — Feature Engineering: Parse .dem → extract 14 metrics per (player, round, side) tuple
2. G2 — Clustering: Ensemble vote → cluster profiles → role assignment → player-level aggregation

---

Pipeline Architecture

N × .dem files  (binary, ~150 MB each)
       │
       ▼  demoparser2 (Rust-based)
       │
  kills / damage / positions / grenades / round_ends
       │
       ▼  reconstruct_rounds()  +  assign_round_num()
       │
  Events tagged with round_num
       │
       ▼  extract_features_from_dem()  ×  N
       │
  14-feature DataFrame  (~3,600 rows for 12 demos)
  ├── g1_features.parquet          [raw]
       │
       ▼  clip → log1p → StandardScaler
       │
  Scaled DataFrame  (+ 14 "_scaled" columns)
  ├── g1_clean.parquet
       │
       ▼  KMeans  +  GMM  +  HDBSCAN  →  majority vote
       │
  Clustered DataFrame  (+ cluster_id, gmm_max_prob)
  ├── g2_clustered.parquet
       │
       ▼  weighted scoring matrix  →  greedy role assignment
       │
  Labeled DataFrame  (+ role_label)
  ├── g2_labeled.parquet
       │
       ▼  groupby player → majority role → consistency threshold
       │
  Player-level summary  (~30 rows)
  └── g2_player_roles.parquet

---

14 Behavioural Features

Dimension	Feature	Range	What It Captures
Combat	kill_count	0–5	Kills per round (capped at ace)
	damage_dealt	0–500	HP damage dealt
	headshot_pct	0–1	HS fraction — riflers ~60%, AWPers ~15%
	multi_kill_round	{0,1}	Binary: ≥3 kills in one round
	opening_duel_won	{−1,0,1}	Won first fight (+1), died first (−1), uninvolved (0)
Utility & Teamplay	utility_used	0–5	Grenades thrown
	flash_assists	0–5	Kills where this player's flash blinded the victim
	trade_kills	0–5	Kills within 5 sec (320 ticks) of a teammate's death
Positioning	distance_traveled	0–5000	2D path length per round
	mean_isolation	0–2000	Mean nearest-teammate distance (cKDTree, same-team only)
	zone_primary	{1,2,3}	Map zone: CT-side / mid / T-side — percentile-adaptive per map
Timing	first_engagement_rel	0–1	Relative tick of first kill (0 = round start, 1 = never)
	death_tick_rel	0–1	Relative tick of death (1.0 = survived the round)
	survived	{0,1}	Binary: alive at round end

✅ G1 Validation: All 14 features pass range checks — confirmed in the validation report printed at end of Cell 5.

---

Clustering Methodology

K Selection

Both metrics are evaluated over K = 3–7; consensus rule picks the final K.

Metric	What It Favours	Current Result
Silhouette Score	Tightest, best-separated clusters	Best K = 3 (score ≈ 0.324)
GMM BIC	Best fit-to-complexity ratio	Best K = 7 (BIC still falling)
Chosen K	Midpoint when disagreement > 1	K = 5 ( (3+7)/2 = 5 )

Ensemble Vote

Three algorithms label every data point independently. Labels are Hungarian-aligned to KMeans before voting. Each point's final cluster is decided by majority.

Algorithm	Configuration	Strengths
KMeans	K=5, n_init=10	Fast, stable; reference for alignment
GMM	Full covariance, n_init=5	Handles elliptical clusters; soft probabilities
HDBSCAN	min_cluster_size=20, min_samples=5	Finds non-convex shapes; noise detection

HDBSCAN noise points (label = −1) inherit the KMeans label before alignment.

Ensemble Agreement (all 3 agree): 38.7%

---

Results — Mid-Term Evaluation (v3)

Cluster Profiles

Cluster	Assigned Role	Rows	%	Key Signals
C0	Support	444	21.9%	Low kills/dmg, high utility & flash assists
C1	IGL	568	28.0%	Low kills, high survival & isolation
C2	Lurker	712	35.1%	High distance & isolation, moderate dmg, late entry
C3	Rifler	145	7.1%	Highest dmg (~400), ~3.8 kills/rd, high HS%, multi-kills
C4	AWPer	159	7.8%	Low kills, very low HS%, high distance, negative opening duel

Player-Level Summary (31 players, Flex threshold = 50%)

Role	Notable Players	Consistency
Lurker	hyped, JDC, LNZ	73%, 63%, 59%
IGL	SunPayus, xfl0ud	50%, 50%
Rifler	zweih	100% (4 rounds only)
Flex	ropz, chopper, tN1R	42%, 43%, 47%

23 of 31 players (74%) fall below the 50% consistency threshold → classified as Flex.
This is the primary open problem — see Issues below.

Visualisation Outputs

Sample outputs generated by the pipeline (see outputs/sample/):

File	Contents
g2_correlation_matrix.png	14×14 Pearson correlation heatmap
g2_k_selection.png	Silhouette score + GMM BIC vs K
g2_pca_tsne.png	PCA and t-SNE cluster views
g2_profile_heatmap.png	Cluster × feature normalised mean heatmap
g2_radar_charts.png	One radar chart per cluster
g2_roles_final.png	Role distribution bar chart + PCA with role colours
g2_player_roles.png	Player consistency histogram + per-player role distribution

---

Bug Fixes: v2 → v3

Fix	Problem in v2	Resolution in v3
FIX 1	trade_kills had no cap — values reached 17 due to double-counting	Added CAP_TRADE_KILLS = 5 and break after first matching death
FIX 2	zone_primary always returned 2 — compute_zone_adaptive([mean_x]) triggered the len < 10 guard	Pass global X distribution (all_x, ~23k values) as reference
FIX 3	Flash assists checked attacker_name + flash_duration > 0 — measured kills-while-flashed, not assists given	Use assister_name == player AND assistedflash == True
FIX 4	No Rifler role — generic fragger clusters mislabelled as AWPer	Added Rifler to scoring matrix with high headshot_pct weight

---

Known Issues & Roadmap

Issue	Severity	Plan
74% players classified Flex — insufficient per-player data	High	Scale to 100+ demos
Cluster imbalance: Lurker = 35%, Rifler = 7%	Medium	More diverse match data
Some cluster profile features show exactly 0 (possible column selection bug)	Medium	Investigate raw vs scaled column indexing
distance_traveled & mean_isolation highly correlated (r = 0.93)	Low	Consider dropping one or using PCA pre-clustering
Entry Fragger role never appears as assigned cluster	Low	Revisit scoring matrix weights

---

Setup

Requirements

• Python 3.9+
• Google Colab (recommended) — the pipeline mounts Google Drive automatically
• CS2 .dem files placed in the configured raw_dems/ folder

Install

pip install -r requirements.txt

Configure paths

Edit the constants in src/cs2_pipeline_v3.py (Cell 1):

_BASE      = "/content/drive/MyDrive/cs2-btp"   # your Drive path
DEM_FOLDER = os.path.join(_BASE, "raw_dems")
OUTPUT_DIR = Path(os.path.join(_BASE, "processed"))

Run (Google Colab)

Execute cells in order — each cell is self-contained:

Cell 1: Setup & configuration
Cell 2: DEM parsing utilities
Cell 3: Round reconstruction & feature helpers
Cell 4: Feature extraction → g1_features.parquet
Cell 5: Preprocessing & scaling → g1_clean.parquet
Cell 6: G2 clustering & role assignment → g2_*.parquet + PNGs

Getting .dem files

Professional match demos are publicly downloadable from HLTV.org under each match's "Demo" tab. Place them in your configured raw_dems/ folder. See data/README.md for details.

---

Repository Structure

cs2-role-classifier/
├── README.md
├── LICENSE
├── .gitignore
├── requirements.txt
│
├── src/
│   └── cs2_pipeline_v3.py        # Full pipeline (Cells 1–6)
│
├── docs/
│   ├── code_explanation.pdf      # Line-by-line explanation of all pipeline stages
│   ├── btp_midterm_slides.pdf    # Mid-term evaluation presentation
│   └── CONTEXT.md                # Design decisions & pipeline context
│
├── outputs/
│   └── sample/                   # Sample visualisation outputs from a test run
│       ├── g2_correlation_matrix.png
│       ├── g2_k_selection.png
│       ├── g2_pca_tsne.png
│       ├── g2_profile_heatmap.png
│       ├── g2_radar_charts.png
│       ├── g2_roles_final.png
│       └── g2_player_roles.png
│
└── data/
    └── README.md                 # How to obtain .dem files (not tracked in git)

---

Notable Correlations (Feature Analysis)

From the correlation matrix generated by the pipeline:

Correlation	Value	Interpretation
distance_traveled ↔ mean_isolation	r = 0.93	High collinearity — both log-scaled but kept
kill_count ↔ multi_kill_round	r = 0.75	Expected — kill count drives multi-kill flag
kill_count ↔ trade_kills	r = 0.71	Players with kills also get more trades
damage_dealt ↔ first_engagement_rel	r = −0.47	Early engagers deal more damage
opening_duel_won ↔ first_engagement_rel	r = −0.48	Winning the entry = engaging early
zone_primary ↔ everything	r ≈ 0	Clean orthogonal feature

---

References

• Kuzhii, Y., & Furgala, Y. (2024). Feature Engineering for Role Assessment in Counter-Strike 2.
• HLTV Rating 2.0: https://www.hltv.org/news/20695/introducing-rating-20
• demoparser2 (Rust-based CS2 parser): https://github.com/LaihoE/demoparser
• HDBSCAN library: https://hdbscan.readthedocs.io

---

Sachin Kumar — UG-2 — S20240010206