SteamAnalysis

Introduction

In the entertainment industry, the gaming sector has established itself as a prominent player (Newzoo, 2023). As game development and releases grow, it has become increasingly crucial for developers and publishers to comprehend the factors and decisions that contribute to a game's triumph. The present study endeavors to explore the feasibility of classifying games into distinct success levels based on their number of owners, positive ratings, and average playtime as features. These factors are believed to be strong indicators of a game's popularity and overall success. Other than quantifying measurements of success, the dataset contains several potential features that may contribute to high ownership which will be explored. Steam is the eminent online software distribution platform. Created by Valve, Steam has become synonymous with PC gaming, which is where the main dataset used is collected.

Dataset

The dataset selected is from Kaggle's Nik Davis Steam Store Games. Collected in May 2019, any game releases or additional data past this date is not reflected in the dataset. The following table describes the features within steam.csv.

appid	name	release_date	english	developer	publisher	platforms	required_age	categories	genres	steamspy_tags	achievements	positive_ratings	negative_ratings	average_playtime	median_playtime	owners	price
Unique ID for each app	Title of app (game)	Release date in format YYYY-MM-DD	Language support: 1 if is in English	Name (or names) of developer(s). Semicolon delimited if multiple	Name (or names) of publisher(s). Semicolon delimited if multiple	Semicolon delimited list of supported platforms. At most includes: windows;mac;linux	Minimum required age according to PEGI UK standards. Many with 0 are unrated or unsupplied.	Semicolon delimited list of game categories, e.g. single-player;multi-player	Semicolon delimited list of game genres, e.g. action;adventure	Semicolon delimited list of top steamspy game tags, similar to genres but community voted, e.g. action;adventure	Number of in-games achievements, if any	Number of positive ratings, from SteamSpy	Number of negative ratings, from SteamSpy	Average user playtime, from SteamSpy	Median user playtime, from SteamSpy	Estimated number of owners. Contains lower and upper bound e.g. 20000-50000.	Current full price of title in GBP, (pounds sterling)

From initial assumptions, features such as 'appid', 'name' or 'achievements' may not have high (or any) feature importance. The main target variable will be 'owners', and since it given as a range, the midpoint of this value will be used.

After some preliminary exploratory data analysis, here are some quick infographs.

The number of native English games (98.11%) to non-English (1.89%)

About 9.46% are free games, 90.54% are paid games.

The number of games released on Steam has exploded since 2014. (Mind the cut off for 2019 limited to the dataset collection)

Most certainly, there are correlations between 'positive_ratings', 'negative_ratings', 'average_playtime', 'median_playtime', and 'owners'. However the more curious finding will be what features that might correlate with success, will a type of category, platform, genre or tag succeed more than others.

There are an abundance of games released on the Steam platform, and as a side effect, great games with a lot of press that receives the most attention will shoot some games right out of the waters. Games such as Dota2, Team Fortress 2, Counter-Strike:Global Offensive are (or have since become) free-to-play, accessible on Mac and Linux, and have dominant Esports presence, notably Dota2 having one of the hiest overall prize pools in Esports history at $40 million dollars (The International 2021, Dota 2 International Esports Tournament), it won't come as a surprise the amount of attention glued to these games are likely to be popular and to have high ownership, resulting in an over successful game that is likely to be considered an outlier.

Since the initial features like 'categories' and 'steamspy_tags' are semicolon delimited, feature engineering techniques like OneHotEncoder and LabelEncoder has been applied to allow some algorithms to be applied.

NLP Results

In the interest of keeping the game title private, the following node graph displays a game with the keywords ‘audio’, ‘price’, and ‘graphics’. While some optimizations based on the removal of non-technical words may be in order, it has proven challenging as some games may contain a variety of different ‘made up’ words, or in-game character names in reference to their review topic. There have been some attempts for the removal of troll comments or dissenters, however it is due to the passion of these gamers through the guise of their anonymity behind a Steam Username that allows for their verbosity, forthrightness and candor, leading to a more honest review of Steam games.

From these node graphs, we can see that with some keywords selected, this game title was not well received, citing gameplay-related issues, bad writing and cringey story, terribly overpriced, and ‘stupid’ graphics.

While the aim of this study is not to replace the face of Steam Reviews, I believe it will give developers/publishers and potential gamers a better understanding of the game in their current state. As is with the human language and the human experience, language and its people will continue to evolve which will require continued analysis and further different techniques to narrow down the fine details of the quality of a game based on text on an online digital gaming platform.

Curious Notes:

In a preliminary exploratory analysis, there are findings that some titles are disproportionately played by certain demographics, while the propensity and willingness of providing a Steam game review does not immediately indicate the proportion of any particular nationality, I believe it remains a strong indicator of the expected player base, especially for online competitive multiplayer games such as Counter-strike.

The following are some early exploratory graphs testing for accuracy and sentimentality. Noting that most words are stopwords that are connecting other related features. There were attempts to create an LDAvis but due to technical issues and time constraints, was unable to procure an image before it crashed. Attempts to create a summarizer using gensim was unsuccessful due to the feature being depriciated in later updates.