The central idea: Dota 2 matches are too complex for a single statistic or model to explain. This project operates as an atlas, not an oracle, combining predictive modeling, draft archetypes, hero relationships, and public-attention patterns to reveal different parts of the competitive landscape.

10,000 Ranked All Pick Matches

90 Dataset Columns

524 Weekly Trends Records

13 Approved Visuals

Executive Summary

The Dota 2 Intelligence Atlas examines a deceptively simple question: how much can the structure of a draft explain about match outcomes, and what additional insight appears when match data is connected to hero relationships and public attention? The project combines 10,000 ranked All Pick matches with Google Trends summaries covering U.S. states, countries, weekly search activity, and five editions of The International. Rather than forcing every question into one predictive model, the analysis proceeds through four complementary layers: outcome modeling, match-profile clustering, hero-pair and network analysis, and geographic and temporal attention mapping.

The predictive stage established the limits of broad composition data. Radiant won 54.45% of the sampled matches, so a model that always predicted Radiant already achieved a meaningful baseline. A logistic regression using 15 draft-composition predictors reached 55.0% accuracy on a reproducible 2,000-match test set, an improvement of only 0.55 percentage points. The confusion matrix showed that the model strongly favored Radiant and failed to identify most Dire victories. This result is not presented as a failed model. It is evidence that broad counts of melee heroes, primary attributes, and roles capture only a small portion of the strategic and executional complexity that determines a Dota 2 match.

Clustering provided a more useful descriptive view. Three recurring match profiles emerged: Frontline Initiation Edge, Agility Carry Edge, and Intelligence Support Edge. Radiant performed best in the Frontline Initiation Edge profile, winning 56.4% of those matches. The profiles overlapped substantially in the principal-component display, showing that drafts exist on a continuum rather than in rigid categories. The clustering results suggest that coherent team identity may be more informative than any single composition variable.

Hero-level relationships produced the clearest actionable signals. The data was expanded into 200,000 same-team pair appearances representing 7,900 unique hero combinations. Frequency and performance were not the same: popular pairs were not automatically the strongest performers. Among established non-Pudge pairs with at least 50 appearances, Spirit Breaker with Wraith King led the displayed ranking with 53 wins in 74 appearances, or 71.6%. A separate co-selection network identified 23 high-frequency connections among 15 heroes, revealing the recurring structural backbone of common drafts without confusing popularity with win rate.

The attention analysis showed that Dota 2 interest is both geographically concentrated and event-sensitive. Wyoming led the U.S. state index, while Mongolia recorded the highest global relative-interest score. These values measure search concentration, not raw audience size. Weekly trends also showed recurring attention spikes around The International. Average tournament-period interest exceeded the corresponding pre-event level in all five observed editions, with lifts ranging from about 41% to 168%.

Taken together, the findings support the central conclusion of the project: Dota 2 cannot be reduced to one decisive statistic. The strongest intelligence comes from combining imperfect but complementary views. The atlas does not claim causal proof or a universal drafting formula. It provides a reproducible, audience-friendly framework for understanding where broad prediction fails, where structural patterns emerge, which hero relationships deserve closer investigation, and when public attention intensifies.

Mission Brief

Research Objective

This project investigates three connected questions:

  1. Can broad draft composition predict which team wins?
  2. Which match profiles and hero partnerships appear most favorable?
  3. Where and when does public attention toward Dota 2 intensify?

The analysis joins a 10,000-match ranked dataset with independently collected Google Trends information covering approximately 5 years.

Analytical Route

The atlas moves through four layers:

Match outcomes -> predictive modeling -> draft structures and hero relationships -> geographic and temporal attention

Each visual is followed by one expandable interpretation containing:

  • Plain language: what a general audience should notice.
  • Deeper interpretation: what the result means analytically and what it does not prove.

The predictive model is intentionally treated as a diagnostic rather than a victory lap. Its modest performance helps reveal that broad team-composition summaries capture only a small part of what determines a Dota 2 match.

Part I - Can Draft Composition Predict Victory?

The Starting Battlefield

Distribution of Radiant and Dire match outcomes
Open interpretation

Plain language

The matches are reasonably balanced, although Radiant wins somewhat more often than Dire. This creates a useful but demanding prediction problem: a model must outperform simply guessing Radiant every time.

Deeper interpretation

Radiant won 5,439 of the 10,000 sampled matches, producing a majority-class baseline accuracy of 54.45%. Any predictive model should therefore be judged against that baseline rather than against an arbitrary 50% benchmark. The imbalance is modest enough to preserve both outcome classes, but large enough that raw accuracy can appear more impressive than it really is.

Baseline Versus Logistic Regression

Comparison of majority-class baseline and logistic-regression accuracy
Open interpretation

Plain language

The logistic model technically performed better than always predicting Radiant, but only by a very small amount. Its 55.0% accuracy exceeded the 54.45% baseline by approximately 0.55 percentage points.

That narrow improvement is itself an important result: broad draft summaries alone do not reliably explain who wins a Dota 2 match.

Deeper interpretation

The model was evaluated on a reproducible 2,000-match test set after training on the remaining 8,000 matches. Because Radiant was the majority outcome, a useful model needed to outperform the 54.45% majority-class baseline.

The logistic model reached 55.0% accuracy. This represents measurable but weak predictive value. The result suggests that attributes such as melee-versus-ranged balance, primary attributes, and broad role composition contain some signal, but much of match outcome depends on information not represented by these aggregate variables.

Important omitted factors may include player skill, hero-specific counters, lane execution, item timing, coordination, objectives, and decisions made during the match.

What the Logistic Model Learned

Estimated logistic-regression feature effects
Open interpretation

Plain language

The model found directional relationships between several draft characteristics and Radiant victory, but none produced a decisive advantage on its own.

A positive estimate indicates that higher values were associated with greater modeled odds of a Radiant win. A negative estimate indicates an association with lower modeled odds. These are tendencies across the sample, not guaranteed effects in an individual match.

Deeper interpretation

The feature-effect chart shows estimated logistic coefficients after removing an exactly redundant predictor from the original specification. That correction left 15 usable predictors and eliminated aliased coefficients.

Coefficient size communicates direction and relative modeled influence while holding the other included variables constant. It should not be interpreted as a causal effect. The modest overall model performance also means that even visually prominent coefficients should be treated cautiously.

The broader lesson is that composition variables help describe a draft, but they do not fully encode matchup quality. Two teams with similar role and attribute counts may contain entirely different hero synergies, counters, lane plans, and execution requirements.

Where the Predictions Broke Down

Logistic-regression confusion matrix
Open interpretation

Plain language

The model strongly favored predicting Radiant. It correctly identified 950 Radiant wins but correctly identified only 150 Dire wins.

Most actual Dire victories were classified as Radiant victories. That imbalance explains why the model’s overall accuracy barely exceeded the simple Radiant-only baseline.

Deeper interpretation

Of the 911 actual Dire wins in the test set, 761 were misclassified as Radiant and only 150 were correctly classified as Dire. This corresponds to an 83.5% Dire misclassification rate.

Among the 1,089 actual Radiant wins, the model correctly classified 950 and misclassified 139 as Dire, correctly recognizing 87.2% of Radiant wins.

In total, the model predicted Radiant 1,711 times and Dire only 289 times. The confusion matrix therefore reveals a model leaning heavily toward the majority class rather than learning a balanced separation between outcomes.

This is why accuracy must be interpreted alongside class-specific performance. A model can produce a superficially acceptable accuracy while remaining ineffective at detecting the less frequently predicted class.

Part I finding: Broad composition measures provide limited predictive signal. The weak model is not a failed analysis; it redirects the investigation toward richer structures such as draft archetypes, specific hero partnerships, and network relationships.

Part II - Draft Archetypes and Match Profiles

The weak logistic model suggested that isolated draft-count variables were not enough. The next stage groups matches by their broader composition patterns to identify recurring draft archetypes.

Three Recurring Draft Profiles

Radiant win rate across three match profile clusters
Open interpretation

Plain language

The clustering analysis identified three recurring match profiles:

  1. Frontline Initiation Edge
  2. Agility Carry Edge
  3. Intelligence Support Edge

Radiant performed best in the Frontline Initiation Edge group, winning approximately 56.4% of those matches. The Agility Carry Edge group followed at 53.9%, while the Intelligence Support Edge group produced a 52.9% Radiant win rate.

These differences are meaningful as descriptive patterns, but none represents an automatic winning formula.

Deeper interpretation

K-means clustering was applied to ten standardized match-profile variables describing rank bracket, melee balance, primary-attribute balance, role balance, and overall draft asymmetry.

A three-cluster solution was selected after comparing total within-cluster variation across candidate solutions from two through six clusters. Additional clusters continued to improve mathematical fit, but the gains diminished and the three-cluster solution remained readily interpretable.

The Frontline Initiation Edge cluster contained 3,255 matches and generally featured more melee heroes, Strength heroes, disablers, and initiators on the Radiant side. Its higher win rate suggests that coherent frontline and initiation advantages may be more useful than isolated role counts.

The Agility Carry Edge cluster contained 3,576 matches and leaned toward Agility heroes and carries, with fewer supports, disablers, and initiators. The Intelligence Support Edge cluster contained 3,169 matches and emphasized Intelligence heroes and supports while reducing melee and initiation presence.

These are associations within observational match data. The clustering process does not establish that selecting one archetype will cause victory.

What Separates the Match Profiles

Principal component map of the three match profile clusters
Open interpretation

Plain language

The cluster map shows that the three profiles overlap substantially. Dota 2 drafts do not divide into perfectly separated categories; instead, the clusters represent recurring tendencies inside a continuous range of possible team compositions.

The colored centers mark the typical location of each profile. Matches farther from a center are less representative of that cluster’s defining pattern.

Deeper interpretation

Principal component analysis was used only to display the ten-dimensional clustering solution in two dimensions. The first principal component explains approximately 23.3% of the standardized variation and the second explains 20.7%, for a combined visual summary of about 43.9%.

Because more than half of the original variation remains outside the two displayed dimensions, overlap in this map should not be interpreted as proof that the clusters are identical.

The map reinforces an important analytical point: draft archetypes are probabilistic profiles rather than rigid boxes. Teams may combine frontline, carry, and support characteristics in many different proportions.

Clustering therefore adds descriptive structure that the logistic model could not provide, but the profiles still operate at a broad composition level. Specific hero combinations may contain more actionable information.

Part II finding: Recurring draft archetypes reveal more interpretable structure than the single predictive model, with the Frontline Initiation Edge profile showing the highest observed Radiant win rate. However, substantial overlap between profiles points toward hero-specific relationships as the next analytical layer.

Part III - Hero Partnerships and Draft Networks

Broad draft archetypes describe the shape of a team. Hero-pair analysis moves closer to the actual draft by examining which specific heroes repeatedly appeared together and how those pairings performed.

Frequency Versus Winning Performance

Interactive view: Density tiles summarize all established pairs from 50 through 200 same-team appearances. Colored circles identify the raw and frequency-adjusted highest and lowest performers.

Open interpretation

Plain language

Frequently drafted hero pairs were not automatically the most successful pairs. Some common combinations produced ordinary win rates, while several less common combinations performed substantially better in this sample.

The chart therefore separates two ideas that are easy to confuse:

  • Popularity: how often two heroes appeared on the same team.
  • Observed performance: how often that team won when the pair appeared.

A pair positioned farther right was observed more often. A pair positioned higher recorded a stronger win rate.

Deeper interpretation

The pair analysis expanded the 10,000 matches into 200,000 same-team hero-pair appearances and identified 7,900 unique pairings. To reduce extreme small-sample volatility, the displayed comparison includes only pairs observed at least 50 times.

That threshold retained 1,042 pairings. It improves stability but does not eliminate selection effects, matchup differences, patch effects, player-skill differences, or the influence of the other three heroes on each team.

The plot shows that draft frequency and win rate capture different dimensions of behavior. High-frequency pairs may reflect familiarity, hero popularity, flexible roles, or established conventions rather than superior performance.

Likewise, a high observed win rate does not prove that a pair independently caused those wins. Pair results should be treated as evidence for further investigation rather than a prescriptive drafting rule.

Highest-Performing Established Pairs

Interactive established-pair analysis: Switch among Established Leaders, Evidence-Adjusted, and Volume vs Win Rate.

The first view ranks non-Pudge combinations with at least 60 co-pick appearances by observed win rate. The second uses the Wilson lower confidence bound to balance strong performance against sample-size uncertainty.

The landscape view places every established pair in the same frequency-versus-performance space. The featured top five remain designated without obscuring the wider pair distribution.

Hover over bars or points for appearances, wins, losses, observed win rate, Wilson-adjusted evidence, and average match duration. These are descriptive associations and do not prove that a hero combination causes victories.

Open interpretation

Plain language

Among pairs with at least 50 appearances, several Wraith King combinations rose to the top of the observed win-rate ranking.

Spirit Breaker with Wraith King led the displayed group at approximately 71.6%, followed closely by Phantom Assassin with Wraith King at approximately 71.2%. Other prominent combinations included Hoodwink with Wraith King, Spirit Breaker with Undying, and Dawnbreaker with Snapfire.

Pudge combinations were excluded from this ranking so that one exceptionally common hero would not dominate the visual story.

Deeper interpretation

The leading non-Pudge pairs and their observed records included:

  • Spirit Breaker + Wraith King: 53 wins in 74 appearances, 71.6%.
  • Phantom Assassin + Wraith King: 47 wins in 66 appearances, 71.2%.
  • Hoodwink + Wraith King: 41 wins in 60 appearances, 68.3%.
  • Spirit Breaker + Undying: 44 wins in 65 appearances, 67.7%.
  • Dawnbreaker + Snapfire: 64 wins in 95 appearances, 67.4%.

These pairings may reflect complementary initiation, durability, control, damage, lane structure, or the wider team environments in which they were selected. The available data supports association, not isolation of a pair-specific causal effect.

The minimum-appearance rule also creates a tradeoff. Raising the threshold would increase statistical stability but remove potentially interesting combinations. Lowering it would retain more pairs while allowing extreme rates driven by very small samples.

The High-Frequency Draft Backbone

Interactive hero draft backbone: Switch among Draft Structure, Win Association, and Evidence Strength while the hero positions remain fixed. Hover over a hero for overall performance and network statistics. Hover near the midpoint of a connection for pair appearances, wins, losses, observed win rate, Wilson-adjusted evidence, and backbone score.

The backbone is intentionally selective: it emphasizes recurring and analytically useful relationships without allowing the most popular heroes to overwhelm the graph. The displayed associations do not establish causation.

Open interpretation

Plain language

The network shows the heroes that formed the most common partnership structure in the sampled matches. Heroes with many visible connections served as recurring hubs in the draft environment.

A connection means that the two heroes appeared together on the same team at least 200 times. It does not mean that the pair had the highest win rate.

Deeper interpretation

The final network uses a 200-appearance threshold, producing 23 high-frequency pair connections among 15 heroes. This threshold intentionally removes the dense outer layer of occasional relationships so the core draft structure remains readable.

The graph is a co-selection network:

  • Nodes represent heroes.
  • Edges represent frequent same-team appearances.
  • Connectivity reflects repeated participation in common pairings.

Network centrality in this context signals draft prevalence and compatibility across many lineups, not causal importance or guaranteed strategic value. A central hero may be popular, flexible, easy to combine with common picks, or widely selected across skill brackets.

The network complements the win-rate charts by distinguishing the draft’s structural backbone from its highest-performing partnerships. Frequency reveals what players repeatedly build; win rate reveals how those observed combinations performed.

Part III finding: Hero-level relationships provide the most actionable layer of the atlas. Popularity and performance are not the same, and several established pairings produced strong observed results. The network adds context by showing which heroes formed the recurring backbone of actual drafts.

Part IV - Mapping Dota 2 Attention

The final analytical layer moves outside the match itself. Google Trends data is used to examine where Dota 2 attention is concentrated and how that attention changes around major competitive events.

Google Trends caution: Every value is a normalized relative-interest index. A score of 100 marks the highest concentration within that specific search pull. It does not represent 100 searches, total players, audience size, or direct comparisons of raw search volume between separately normalized geographies.

Dota 2 Attention Across the United States

Interactive attention map: Hover over any state to inspect its exact relative-interest index, national rank, percentile, and attention tier. These values measure search concentration, not raw audience size or total search volume.

Open interpretation

Plain language

Dota 2 search attention was not evenly distributed across the United States. Wyoming recorded the highest relative concentration in this pull, followed by Washington, California, Alaska, and Utah.

The map emphasizes concentration rather than population. A smaller state can receive a high score when Dota 2 searches make up a comparatively large share of that state’s total search activity.

Deeper interpretation

The state-level data contains one observation for each of the 50 states. Leading relative-interest values included Wyoming at 100, Washington at 82, California at 65, Alaska at 59, Utah at 57, Massachusetts and Nevada at 54, and Kansas at 51.

These values should not be interpreted as estimates of total fans or players. California may generate more total searches than a smaller state while receiving a lower relative-interest score because Google Trends normalizes search concentration within each location.

The four attention tiers provide a descriptive geographic summary. They do not test whether differences between states are statistically significant, nor do they explain why interest is elevated in a particular location.

Global Relative Interest

Interactive global attention atlas: Use the buttons above the map to switch between percentile-based attention tiers and the original Google Trends index. Hover over a country to inspect its exact index, worldwide rank, percentile, and attention tier.

Google Trends measures relative search concentration within each country. These values do not represent total searches, player population, or absolute market size.

Open interpretation

Plain language

The global ranking highlights several countries where Dota 2 occupies an unusually large share of search attention. Mongolia stands out at the indexed maximum of 100, followed by Myanmar, Peru, and the Philippines.

This does not mean Mongolia generated the greatest number of Dota 2 searches worldwide. It means Dota 2 was especially concentrated relative to other searches within the country during the selected period.

Deeper interpretation

The global dataset contains 76 countries. The highest indexed values were Mongolia at 100, Myanmar at 55, Peru at 51, the Philippines at 48, North Macedonia at 28, Cuba at 26, and Bolivia at 25.

Mongolia is a visible outlier: its indexed value is approximately 1.82 times the second-highest country. The distribution is strongly right-skewed, with a median of 4 and a mean of approximately 9.1.

The outlier should be retained rather than automatically removed because it may represent a real concentration of interest. However, it should be interpreted as a normalized search signal rather than evidence of the world’s largest player base or esports market.

Cross-country differences may also reflect language, search-engine usage, internet access, regional esports culture, and the way Google Trends samples and normalizes activity.

Attention Over Time

Interactive timeline: Hover over either series for weekly relative-interest values. Use the range slider or drag directly across the chart to inspect specific periods. The shaded regions mark TI10 through TI14 and their surrounding 30-day windows.

Open interpretation

Plain language

Dota 2 attention moves in bursts rather than remaining constant. The largest spikes frequently occur near The International, the game’s premier annual tournament.

The red bands show the period beginning 30 days before each tournament and ending 30 days afterward. The darker center marks the actual tournament dates.

Deeper interpretation

The time-series data includes 262 weekly worldwide observations and 262 weekly U.S. observations from July 2021 through July 2026.

The worldwide and U.S. lines were produced from separate Google Trends pulls and are independently normalized. Their vertical positions can be compared as within-series attention levels, but they should not be read as direct raw-volume comparisons between the United States and the world.

The event windows reveal recurring temporal alignment between major tournament periods and elevated attention. This pattern supports an event-attention association, but it does not establish that every increase was caused exclusively by The International. Updates, patches, roster news, other tournaments, and broader gaming events may also contribute.

The International Event Pulse

Interactive event pulse: Switch between Attention Profile and Event Lift.

The profile view compares worldwide relative interest during the 30 days before each tournament, throughout the tournament period, and during the 30 days afterward.

The lift view ranks each edition by its percentage increase over its own pre-event baseline. Hover over a bar for the tournament dates, host city, relative-interest values, event duration, and calculated lift.

Google Trends values are normalized relative-interest indices rather than raw search counts or audience totals.

Open interpretation

Plain language

Average search attention increased during every observed edition of The International.

TI10 produced the largest percentage lift relative to its pre-event period at approximately 168%. TI11 increased by about 92%, TI14 by 86%, TI13 by 64%, and TI12 by 41%.

Attention usually fell after the tournament, although TI14 retained a slightly higher post-event average than its pre-event level.

Deeper interpretation

The event-pulse comparison uses three windows for each tournament: the 30 days before the event, the actual tournament dates, and the 30 days after the event.

Average relative-interest values were:

  • TI10: 30.25 before, 81.0 during, and 28.75 after.
  • TI11: 46.75 before, 89.67 during, and 36.0 after.
  • TI12: 39.25 before, 55.33 during, and 34.25 after.
  • TI13: 38.5 before, 63.0 during, and 35.0 after.
  • TI14: 46.5 before, 86.5 during, and 49.0 after.

The consistent tournament-period lift provides strong descriptive evidence that The International functions as a recurring attention catalyst. The varying magnitude across years also shows that the size of the effect is not fixed.

Because the analysis is observational and does not control for simultaneous releases, patches, promotions, or news cycles, the result should be described as temporal association rather than causal proof.

Part IV finding: Dota 2 attention is geographically concentrated and strongly event-sensitive. Search interest rises around The International in every observed year, while state and country patterns reveal where the game occupies an unusually large share of local search attention.

Part V - What the Atlas Reveals

The atlas began with a direct prediction question: can broad draft composition explain who wins a Dota 2 match? The answer was only slightly. That limitation became the path to the more useful findings.

Findings Across the Four Analytical Layers

55.0% Logistic Test Accuracy

56.4% Top Cluster Radiant Win Rate

71.6% Top Established Pair Win Rate

5 of 5 TI Events With Attention Lift

1. Prediction exposed the limits of broad composition

The logistic model reached 55.0% accuracy, only 0.55 percentage points above the 54.45% majority-class baseline. It also leaned heavily toward predicting Radiant and performed poorly when identifying Dire victories.

The model therefore captured a small amount of structure, but not enough to function as a reliable match predictor.

2. Clustering revealed recurring draft identities

The three-cluster solution produced interpretable match profiles centered on frontline initiation, agility carries, and intelligence supports.

The Frontline Initiation Edge profile recorded the highest Radiant win rate at approximately 56.4%. The result suggests that draft coherence may matter more than any individual hero-count variable viewed in isolation.

3. Hero partnerships produced the clearest actionable signals

Pair frequency and pair performance were not interchangeable. Some of the most common combinations delivered ordinary results, while several established but less frequent pairs recorded much stronger observed win rates.

The strongest displayed non-Pudge pairing was Spirit Breaker with Wraith King, winning 53 of 74 appearances for an observed rate of approximately 71.6%.

The network analysis added context by identifying 15 heroes and 23 high-frequency relationships that formed the draft environment’s recurring backbone.

4. Public attention was geographic and event-driven

Google Trends revealed substantial geographic concentration in both the United States and the global sample. These signals represent relative search intensity rather than raw audience size.

The International also produced an attention increase in every observed year. Tournament-period interest rose between approximately 41% and 168% above the corresponding pre-event average.

Central conclusion: The most useful Dota 2 intelligence did not come from asking one model to predict everything. It came from combining several imperfect views of the same ecosystem.

Data and Methods

The project integrates two analytical domains:

Domain Data Primary Methods Purpose
Match intelligence 10,000 ranked All Pick matches with 90 source and engineered columns Descriptive analysis, logistic regression, confusion-matrix evaluation, standardization, k-means clustering, principal component analysis, hero-pair aggregation, and network analysis Examine outcomes, draft profiles, hero partnerships, and recurring co-selection structures
Public attention U.S. state, global country, weekly time-series, and TI event-window Google Trends summaries Geographic tiering, indexed ranking, longitudinal comparison, and event-window analysis Identify where and when relative Dota 2 attention intensifies

The match sample was divided reproducibly into 8,000 training observations and 2,000 testing observations for logistic-model evaluation. The clustering analysis used ten standardized match-profile features and a three-cluster solution selected for interpretability and diminishing improvements in within-cluster variation.

Hero-pair analysis expanded the matches into 200,000 same-team pair appearances. The performance comparison required at least 50 appearances, while the network retained relationships observed at least 200 times.

What the Analysis Does Not Prove

Several limitations shape the interpretation:

  • The match data is observational and cannot establish that a draft characteristic or hero pair caused a victory.
  • Aggregate composition variables omit player skill, lanes, items, counters, objectives, communication, execution, patch conditions, and decisions made after the draft.
  • Pair win rates include the influence of the remaining heroes, opponents, ranks, and match contexts.
  • Cluster labels are analytical summaries assigned after examining standardized group characteristics.
  • Google Trends values are normalized indices, not raw searches, player counts, market sizes, or direct measures of fandom.
  • Tournament attention may overlap with patches, announcements, promotions, roster changes, and other esports activity.

These limits do not invalidate the findings. They define the boundary between what the evidence supports and what would require richer data or experimental design.

Practical Interpretation

For analysts, the project demonstrates why model performance should be compared against a meaningful baseline and examined with class-specific diagnostics.

For players and draft researchers, the results suggest that specific hero relationships and coherent team identities may contain more information than simple counts of attributes and roles.

For marketers, tournament organizers, and community researchers, the Trends analysis shows that attention is concentrated geographically and repeatedly activated by major competitive events.

Next Analytical Frontier

A stronger future system could add:

  • hero-specific matchup and counter variables;
  • lane assignments and role positions;
  • player or team skill controls;
  • patch and version identifiers;
  • item timing and objective progression;
  • nonlinear models such as random forests or gradient boosting;
  • network centrality and community-detection measures;
  • uncertainty intervals for pair performance;
  • additional event and media signals surrounding attention spikes.

These additions would move the project from broad match intelligence toward a more complete drafting and outcome system.

Final takeaway: Dota 2 is not governed by one decisive variable. Victory emerges from interacting heroes, team structures, player execution, and match context. Public attention behaves similarly: it concentrates in particular places and surges around major moments. This report is therefore best understood as an intelligence atlas—a collection of complementary maps that make a complex competitive ecosystem easier to navigate.

Responsible AI Use Statement

OpenAI ChatGPT was used during this project as a development, debugging, design, and writing assistant. AI-supported tasks included reviewing R code, identifying possible workflow errors, suggesting visual and report structures, improving code comments, and helping translate technical findings into plain-language explanations.

The author retained responsibility for the research questions, dataset selection, feature definitions, analytical thresholds, model choices, interpretation, and final presentation. AI-generated code was not accepted as correct without execution and review in the project’s R environment. Numerical findings included in the report were checked against the underlying datasets, model outputs, tables, and approved visualizations.

Responsible-use safeguards included:

  • Executing and testing suggested code before incorporating it into the final workflow.
  • Checking reported values against source data rather than relying on generated prose.
  • Rejecting or revising suggestions that produced incorrect, unclear, or misleading outputs.
  • Separating observed association from causal claims.
  • Preserving limitations, uncertainty, and weak model performance rather than overstating results.
  • Using only the project’s actual match and Google Trends data for analysis.

Generative AI was not used to fabricate match records, search-trend observations, statistical results, or citations. It also did not independently determine which findings should be accepted. The final analysis represents a human-reviewed workflow in which AI served as a collaborative tool rather than an autonomous analyst. Responsibility for the accuracy, interpretation, and academic integrity of the final submission remains with the author.

Reproducibility and Deliverables

The final report uses:

  • one 10,000-match analytical dataset;
  • four Google Trends summary datasets;
  • thirteen approved static analytical visuals;
  • root-relative image paths;
  • a self-contained HTML output suitable for RPubs publication;
  • expandable interpretations separating plain-language explanation from deeper analytical context.

All report visuals are loaded from verified PNG files in the project root. Knitting the report does not regenerate or alter the approved graphics.

Dota 2 Intelligence Atlas: Predictive Modeling, Hero Draft Networks, and Google Trends Mapping