Analysis of Users’ Sentiment towards Competing Video Gaming Consoles on Twitter
Oghenerume Efeduma
10 October, 2023
Introduction
The goal of this analysis is to understudy and compare user’s sentiments towards the two largest game consoles globally - Playstation 5 from SONY and the Xbox Series X from Microsoft and offer insights that could be helpful to various stakeholders especially in the context of product marketing and product development. A worthy third competitor, the Nintendo DS is excluded from this analysis as it focuses only on non hand-held consoles.
Methodology :
- Tweets sourced using the twitter REST API and rtweet package [1]
- Exploratory data analysis, data cleaning and data transformation
- Sentiment analysis and statistical inferences
- Development of ranking algorithms for most influential twitter users that talk about both products.
This analysis provides a replicable template for anaylzing twitter users’ sentiments towards any two competing products
Skills: Sentiment Analysis, Exploratory data analysis, data mining, ranking algorithms, data visualization, statistical inference
Why ‘PlayStation 5’ vs ‘Xbox Series X’?
Microsoft and SONY are two of the biggest players in the non hand-held console gaming market with their feature products ‘Xbox’ & ‘PlayStation’ being 2 of the most purchased gaming consoles globally in the last 2 decades.[2]. In November 2020, Microsoft released the “Xbox Series X”(Xbox) to rival SONY’s “Playstation 5”(Ps5) released in same month.[3]. These global sales figures up until 2021 are presented below.
Who needs to understand User Perception of these Products?
Some key stakeholders that could benefit from this analysis include:
- International retailers who need to understand which product will improve their sales outcomes (Amazon, Walmart etc.)
- Video game developers who need to understand what console to channel their game development efforts (Activision, Gameloft etc.)
- The product manufacturers (Microsoft and SONY) who would like to know how their performing against the competition.
Data Gathering
The tweets used in this analysis were sourced Data sourced using the ‘rtweet’ package in R. [1]. 16470 tweets were downloaded for Xbox Series X and 34567 tweets downloaded for PlayStation 5. The Rtweet package limits tweet search to within 6-9 days leading up to the date of download which in this case was November 24th 2022. [1]. English speaking markets were the focus of analysis and this was specified for the tweet search. Retweets were excluded.
A Preliminary Study to understand how users mention these products in their tweets was used to structure the keyword search which included multiple keywords.
Xbox Series X search keywords: “xbox series x”,“xboxseriesx” or #xboxseriesx
Playstation 5 search keywords: “ps5”, “playstation 5”,“playstation5”, #playstation5 or #ps5
Data Structure
The data consists of 51037 rows/tweets and 91 columns/variables with several text, boolean and numeric variables. The key variables of the tweets used in this analysis include:
user_id[text]: Unique id of tweet authorfollowers_count[numeric]:Count of followers of tweet authorverified[Boolean]: Verification status of tweet authorscreen_name[text]: Handle of tweet author
created_at[datetime]: Time of tweet creationtext[text]: The tweet textfavorite_count[numeric]: Number of favorites of the tweet
retweet_count[numeric]: Number of retweets of the tweetplace_url[text]: Link to the geographic location data of userlocation[text]: User-specified location
Data Cleaning and Transformation
Using the visdat package in R, I quickly analyzed the columns of interest to understand the consistency in datatypes and missing values in the data.
On analyzing the selected tweet features/variables the location
related data (place_url,location), were most
inconsistent with multiple empty entries, inconsistent formatting and
weird values. I further analyzed the location related data to check for
reliability and only 1% of tweets had geographical coordinates and 59%
had user-specified locations. 40% of all tweets had no location
data.
| Tweet Count | % of Total Tweets | |
|---|---|---|
| Tweets with geographical coordinates | 561 | 1% |
| Tweets with user-specified Locations | 30348 | 59% |
| Note: | ||
| This table considers a total of 51,037 tweets downloaded for both products in the analysis |
Looking into user-specified locations I observed inconsistencies with the location formatting and lots of potentially misleading values like “NA”, worldwide” etc.
| Location | Tweet Count |
|---|---|
| NA | 20689 |
| Pittsburgh, PA | 1964 |
| United States | 1860 |
| United Kingdom | 352 |
| Worldwide | 336 |
| Note: | |
| This table considers a total of 51,037 tweets downloaded for both products in the analysis |
From the preliminary analysis it is clear that analyzing metrics by
location from this data will not be very reliable. Nonetheless, I
attempted to manually map several of the readable user-specified
location values to associated countries and combine with the more
reliable geo_location data from place_url. The tweets
successfully mapped only accounted for approximately 30% of the data and
is distributed as below
Product Popularity Analysis
Is the ‘PlayStation 5’ more tweeted about than the ‘Xbox Series X’ from our data?
The popularity of a subject or product on Twitter can be quantified by how much a Product is talked about. This can be measured specifically in tweet frequency over time which forms the basis of of features like “Trends” on Twitter.Due to more frequent tweeting related to the ps5 during the period considered, twice as many tweets were downloaded about the ps5 than the xbox and within a much shorter time frame.
For a more direct comparison of tweet frequencies I computed the average
tweet frequency in tweets per minute (TPM) for the tweets related to
each product be creating a new column called minute_created
(by reducing the datetime value in the created_at column to
the nearest minute and aggregating the number of tweets in each minute
as the tweet_per_minute
## # A tibble: 6 × 3
## # Groups: subject [2]
## subject minute_created tweet_per_minute
## <chr> <dttm> <int>
## 1 ps5 2022-11-21 02:58:00 7
## 2 ps5 2022-11-21 02:59:00 4
## 3 ps5 2022-11-21 03:00:00 18
## 4 xbox 2022-11-15 14:34:00 2
## 5 xbox 2022-11-15 14:41:00 2
## 6 xbox 2022-11-15 14:45:00 1The summary statistics showed a higher mean TPM for the Ps5 versus the Xbox but the standard deviation is pretty high showing a wide dispersion of values. For context, in a normal distribution, 99% of the values occur within a range of 3 times the standard deviation. The max value is well above this threshold, hence, it is very likely that there are outliers in our TPM distribution and a boxplot will highlight this effectively. I removed these outliers to have a more representative distribution of the tweets per minute on each product.
## # A tibble: 2 × 6
## Product `Mean Tweet Frequency` `Standard Deviation` Median Max Min
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 ps5 8.32 4.76 7 38 1
## 2 xbox 2.06 1.41 2 19 1
After removing outliers these are the updated summary statistics:
## # A tibble: 2 × 6
## Product Mean_Tweet_Frequency Standard_Deviation Median Max Min
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 ps5 7.89 4.02 7 19 1
## 2 xbox 1.93 1.11 2 5 1
The density plot below shows the difference in the distribution of TPM for both products. TPM for ps5 tweets tend to be much higher and vary over a wider range than the tweet frequency for the Xbox
How popular is the ‘Playstation 5’ vs ‘Xbox Series X’ in major Markets?
Using the user-specified location data (with the caveat that it isn’t very reliable), I compared the mean TPM in the largest markets/location to give an idea of how the product popularity varies across these markets. The relative product popularity tends to vary by country from the data. For instance in Canada the Ps5 has a higher mean TPM than the Xbox and in USA the reverse is the case. It’s important to note that this data only represents a part of the 30% location data that was mapped earlier. This analysis can be supported by actual product sales data in these countries but this is outside the scope of this study.
| Country | Playstation 5 | Xbox Series X | diff (Ps5-Xbox) |
|---|---|---|---|
| Australia | 1.12 | 1.04 | 0.08 |
| Canada | 1.03 | 1.01 | 0.02 |
| France | 1.07 | 1.04 | 0.03 |
| United Kingdom | 1.13 | 1.22 | -0.09 |
| United States | 1.58 | 1.61 | -0.03 |
| Note: | |||
| values in tweets per minutes |
Word Analysis and Text Mining
As a preliminary step to the lexicon sentiment analysis approach taken in this study, I carried out a word analysis. This included the following steps:
- Breaking down the tweet text into a list of words
- Removing spaces, filler words, stop words/less descriptive words and characters (numbers, symbols, urls)
- Checking for the most frequently used words in the tweets
Unnesting an cleaning tweet text
## # A tibble: 6 × 3
## subject tweetnumber word
## <chr> <int> <chr>
## 1 xbox 1 dungeonsgame
## 2 xbox 1 bringing
## 3 xbox 1 cross
## 4 xbox 1 progression
## 5 xbox 1 game
## 6 xbox 1 pcAnalyzing word frequency
I created a wordcloud of the words used in the tweets to highlight the words that were most recurring in tweets related to each product. - For the Xbox, negative words like hacking and hacked most frequent which indicated a possible negative sentiment towards the product. - For the ps5 we observed frequent mentions of the Playstation4(Ps4), a previous version of the ps5 and on reviewing a sample of tweets, I observed a lot of comparison by users between the ps5 and it’s predecessor.
Word frequency in related tweets
Sentiment Analysis
Is the Twitter User Sentiment better toward the ‘PlayStation 5’ than the ‘Xbox Series X’ from our data set?
The sentiment analysis was carried out as follows:
- I imported the Bing Lexicon, a preset list of words and their associated sentiments (negative or positive)
- Assigned the words from the tweets their sentiments from the lexicon. See below:
# A tibble: 5 × 3
`Tweet Number` Word Sentiment
<int> <chr> <chr>
1 2 lost negative
2 2 recover positive
3 2 recover positive
4 2 discord negative
5 3 loved positive - Calculated the number of positive and negative words per tweet
- Found the sentiment score per tweet as the difference between the positive word count and negative word count for that tweet. A tweet with more positive words will have a more positive sentiment score and vice versa.
# A tibble: 5 × 4
`Tweet Number` negative positive score
<int> <dbl> <dbl> <dbl>
1 10001 0 1 1
2 10002 1 0 -1
3 10004 0 1 1
4 10007 1 0 -1
5 10008 0 6 6The distribution of sentiment scores of the tweets tended towards a normal distribution. The mean sentiment score for the ps5 was higher(more positive) than the Xbox Series X as expected from our word analysis. Nonetheless both mean sentiment scores tended towards 0 a neutral sentiment.
## # A tibble: 2 × 6
## subject mean_score sd Median Max Min
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 ps5 0.116 1.47 1 7 -6
## 2 xbox -0.0570 1.26 0 6 -6
Is the Twitter User Sentiment better toward the ‘PlayStation 5’ than the ‘Xbox Series X’ for the wider population of twitter users?
Considering the sentiment scores are normally distributed from our data, it follows that some simple statistical inferences can be made about the sentiments of the wider twitter population.(Some knowledge of standard statistical testing will be required to fully understand the steps below)
Taking 10,000 Randomly sampled tweets for each product and considering the unequal Variances of Sentiment Scores for Tweets of both products, two statistical tests can be carried out:
- A two-tailed welch t-test to confirm a difference between the mean sentiment scores for both products.
- A one-tailed welch t-test to assert that the sentiment score for the ps5 is indeed higher than the Xbox.
Two tailed-Welch T-test formulation:
\[ \begin{aligned} H_0:& \mu_{Ps5} = \mu_{xbox} \quad \text{: } \mu_{Ps5}\text{ sentiment mean is equal to } \mu_{xbox} \\ H_1:& \mu_{Ps5} \neq \mu_{xbox} \quad \text{: } \mu_{Ps5} \text{ sentiment mean is not equal to } \mu_{xbox} \\ \end{aligned} \] One tailed-Welch T-test formulation:
\[ \begin{aligned} H_0:& \mu_{Ps5} \leq \mu_{xbox} \quad \text{: } \mu_{Ps5}\text{ sentiment mean is less than or equal to } \mu_{xbox} \\ H_1:& \mu_{Ps5} > \mu_{xbox} \quad \text{: } \mu_{Ps5} \text{ sentiment mean is greater than } \mu_{xbox} \\ \end{aligned} \]
Results:
| Test Description | p-value |
|---|---|
| Two-tailed Welch’s test | 4.353735e-10 |
| One-tailed Welch’s test | 2.176867e-10 |
From the results of both tests the p-value is less than 0.05. This means:
The two-tailed t-test disproves the null hypothesis that there is no difference between the mean sentiment scores for both products and allows us to conclude with some certainty that the sentiment scores for both products are not the same for the wider twitter population
The one-tailed t-test disproves the null hypothesis that there the mean sentiment score for the ps5 is less than that of the Xbox. This allows us to conclude with some certainty that the sentiment scores for the ps5 is higher (more positive) than for the Xbox
User engagement and Influencer Analysis
Are Twitter Users engaging more with with ‘Xbox Series X’ or ‘Playstation 5 Tweets’?
By simply trying to answer this question using the mean values of
tweet engagement metrics (retweet_count,
favorites_count) we find similar mean
favorites_per_tweet and retweets_per_tweet for
both products.
Nonetheless, A more detailed look at the median, range and skew of the data tells a very different story that’s closer to reality. The majority of tweets on twitter do not average ten favorites or 2 retweets as the previous chart suggests.In reality most tweets actually get 0 retweets and likes, and there are several factors that cause tweets to garner more engagement. Considering a median retweet_per_tweet and favorite_per_tweet of 0 and high skew, the mean is not reflective of the data. A deeper analysis will be required to understand the factors that drive higher user engagement with these tweets. This is outside the scope of this study.
## n mean sd median trimmed mad min max range skew
## retweet_count 51037 1.70 54.87 0 0.05 0 0 7866 7866 86.69
## favorite_count 51037 10.43 416.74 0 0.50 0 0 70320 70320 125.91Who is driving the engagements with these products on Twitter?
To answer this final question, I developed an algorithm to rank the top twitter accounts (Influencers) driving user engagement around these products which could be of marketing interest. The algorithm considers factors such as tweet frequency, user engagement and author sentiment to rank the top influencers in discussion about these products. The influence rank algorithm also gives more weight to user engagement with the author’s tweets than to metrics controlled by the author like tweet frequency.
| User | Rank | Followers | Tweet count | Mean Retweets per tweet | Mean favorite per tweet | Mean Sentiment Score |
|---|---|---|---|---|---|---|
| Atlus_West | 2907 | 421736 | 3 | 281 | 2136 | 0.33 |
| Wario64 | 2903 | 1155165 | 3 | 29 | 319 | 0.33 |
| JamieMoranUK | 2897 | 23558 | 26 | 6 | 94 | 0.54 |
| IGNDeals | 2893 | 287215 | 3 | 17 | 178 | 1.00 |
| gundamevolution | 2891 | 35672 | 3 | 253 | 1257 | 0.67 |
| User | Rank | Followers | Tweet count | Mean Retweets per tweet | Mean favorite per tweet | Mean Sentiment Score |
|---|---|---|---|---|---|---|
| NextGenPlayer | 9639 | 66869 | 6 | 114 | 899 | 0.17 |
| GermanStrands | 9630 | 36592 | 10 | 81 | 956 | 0.70 |
| videogamedeals | 9599 | 475742 | 10 | 7 | 42 | 0.30 |
| LimitedRunGames | 9568 | 237760 | 6 | 6 | 31 | 0.17 |
| TCMF2 | 9550 | 11917 | 4 | 65 | 558 | 0.50 |
Limitations of the analysis
Twitter API Limitations
- Variables such as reply_count and quote_count were not available to us to improve the analysis of consumer engagement and reach.
- Twitter location data is not reliable.
Limitations of the Statistical tests
- Social media data is not representative of the actual product consumer population.
- Sentiment scores can be skewed by singular events/incidents e.g a likely hacking incident related to Xbox Series X in this case.
- A more effective deployment of this analysis will be to a continuous stream of twitter data instead of tweets for a brief period.
Limitations of Lexicon-based approach
- Studies show that the lexicon based approach not as accurate as supervised-learning approach. ([4]).
- Lexicon-based methods are restricted by their preset lexicons, and static sentiment values of words regardless of the context.This sentiment analysis cannot account for context, sarcasm and other nuances of written communication ([4]).
- In the context of twitter where non-traditional language is largely used with different colloquialisms generated each day, it’s even more difficult for lexicon based approaches to be manually updated to account for new words. ([4]).
Key Insights & Strategic Suggestions
Knowing what product has more traction on twitter can enable International Retailers optimize their marketing strategy for either product.
International Retailers can leverage on influencers for each product as part of their marketing strategy to improve the engagement and sentiment of users around that product.
Video Game Developers can focus their resources on developing more games for the product that garners better positive traction.
Manufacturers could monitor user sentiments for customer feedback that can be channeled to product improvements. Microsoft for instance could look into a possible hacking related incident revealed from the Word Analysis of the Xbox Series X tweets.
Also from the Word Analysis, SONY could look into PlayStation 4(Ps4) related tweets to understand possible preferences for features from of the Ps4 to maintain in version of the ps5.