Meeple

for People

Mohinish (Mo) Shukla

Insight Data Science Fellow

Boston 2020

MEANINGful recommendations for

board gamers

​

1

PERSONAL CONNECTIONS AND INTERACTIONS ARE INCREASINGLY VALUABLE

A GROWING MARKET^a WITH A GREAT DEAL OF COMMUNITY SUPPORT^b

SOCIAL BONDING

AND

CREATIVE EXPRESSION

Why Board Games

2

​

​

a. https://www.reportlinker.com/p05482343/Board-Games-Market-Global-Outlook-and-Forecast.html

b. https://www.tabletopgaming.co.uk/news/games-have-now-made-1-billion-on-kickstarter-and-tabletop-projects/

~$700million on Kickstarter since 2009

Problem:

Finding new games relies on surface features

• Games judged more holistically
• Not well captured by such surface features

3

A novel strategy to augment recommendations

Create a semantic (meaning) space to capture similarity between games

Validate with survey & predicting individual users’ rankings

4

5

Feature data

Semantic data

Scraped ~ 18k games from BoardGameGeek

Descriptions by publishers and editors

Surface features such as type, category, etc.

6

4. Calculated game similarities in semantic space

3.Projected games in semantic space

2. Semantic game space as a 100- feature word2vec embedding

02

01

03

02

1. NLP-ready game descriptions

04

03

2.Projected games in feature space

1.Extracted 127 game features

01

3. Calculated game similarities in feature space

Building models to predict similar games

7

validation

A survey and a novel metric

1. Board game users endorse recommendations made by both models (survey, N=121)

9

Target: Carcassonne.

Which would you say is the closer match to Carcassonne?

​

1. Oklahoma Boomers

(semantic model)

​

2. Fjords

(feature model)

👉 Additionally, users were far more likely to know games with high average scores

2. A user’s top 3 games predicts how they rank the rest of their games (BGG users, N=530)

Model-derived ranks are significantly better than a random model

10

Chance

​

​

Prediction quality ➤

(Mean Spearman’s ρ)

2. A user’s top 3 games predicts how they rank the rest of their games (BGG users, N=530)

Model-derived ranks are significantly better than a random model

11

Chance

​

​

Semantic similarity

+

Featural similarity

+

Avg. Score

normalized +

Popularity

normalized

Prediction quality ➤

(Mean Spearman’s ρ)

/4

MEANINGful recommendations

Summary

Semantics-based recommendations

Improve on traditional content-based recommendations

​

Broad application!

augmenting recommendations with

domain-specific semantics.

Novel validation method!

Rank distribution prediction

​

12

MOHINISH (MO) SHUKLA

13

Molecular genetics, ecology

Neurobiology, Cognitive neuroscience

Infant and adult cognition, neuroimaging, decision making

THANKS

14

CREDITS: This presentation template was created by Slidesgo, including icons by Flaticon, and infographics & images by Freepik

​

​

Game score and popularity

Additional variables to improve recommendations

15

Most

popular

Best

Validation by predicting user game rankings

Each user’s top 3 and bottom 3 rated games were used to predict their rankings of their remaining games

16

Chance

​

​

Prediction quality ➤

(Mean Spearman’s ρ)

Features EDA

“Surface” features (N=127) are sparsely distributed, & only features with low correlations are retained

17