Backing Up Reads (not necessarily game-related, and indirectly backing up reads so I can just read-drop instead of regurgitating reasoning).
I've been doing a bit of research as of late. To be more specific, some (relatively) intense web scraping has been performed on this site, WG, Mafia Universe, and The Syndicate, and especially MafiaScum, over the past year or so in tandem with this dataset (which contains many games obtained from here). I will not be divulging much, depending on what one would define as much (no graphs, not including complete methodology and key data, so a lot of this could be considered by some to be a "whole lot of nothing"). Obviously (as noticed from the past few games) my usage of this data and understanding of my tells hasn't truly been exactly the best. I will formally be putting my training/testing to the test in a game I'm participating in soon, as I believe I have achieved sufficiently good accuracy.
Some things I've done
A Few of My Findings
------
Some of the findings I realized (vast majority obviously not seen here) are built upon the findings of retired mafia player Ellibereth.
None of this is to say my play these past few months has been all of a greater scheme to refine and find better tells to obtain better reads and all, or that I have been trolling in focus of research and not playing to the best of my ability; it is ultimately the result of subpar play, tunneling, and pushes. These last few games only built upon me testing the efficacy and usefulness of common tells on a surface-level.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
1) I believe it decreased hostility as opposed to what it would be if I had responded.
2) In order to add to my pool of data to assess the aggressiveness as an effective scum-tell when applied on the surface (thought process not considered).
3) To read Lethal (especially Lethal, as I scraped many of their posts and noticed their past AI accusations against others (likely not serious), The X, and others with a certain degree of accuracy.
Examining read accuracy (using same method as Ellibereth)
To measure the accuracy of my reads and track improvement, I used the Brier Skill Score (BSS). This method compares the accuracy of my probabilistic reads (e.g., stating Player A is 80% Town) against a baseline "no skill" reference forecast. This "no skill" forecast represents someone guessing alignments based only on the known number of players for each alignment at that point in the game (e.g., if 4 players remain with 3 Town and 1 Mafia, the no-skill guess is 75% Town for everyone).
The calculation involves first finding the Brier Score (BS) for my set of reads. The Brier Score is the mean squared difference between my forecasted probability (ft) and the actual outcome (ot, where 1 means the predicted alignment was correct, and 0 means it was incorrect). The formula is:
BS = (1/N) * Σ(ft - ot)²
Here, N is the total number of reads I'm evaluating. A lower Brier Score is better, with 0 being a perfect score.
Next, I calculated the Brier Score for the "no skill" reference (BSref) over the same set of players/reads, using the base rate probabilities (derived from player counts per alignment) as the forecast (ft) for each read in the formula.
Finally, I computed my Brier Skill Score (BSS) using my Brier Score (BSf) and the reference Brier Score (BSref):
BSS = 1 - (BSf / BSref)
A BSS of 1 means my reads were perfect. A BSS of 0 means my reads were exactly as accurate as the "no skill" guesser. A positive BSS (between 0 and 1) indicates my reads were better than the baseline, while a negative BSS means my reads were worse than simply guessing based on the numbers. While this method works best for probabilistic reads, it can be adapted for binary reads ("Player is Town" or "Player is Mafia") by treating them as 100% certainty (1.0 or 0.0 probability), though this often results in poorer scores compared to nuanced probabilistic reads. This calculation can be done for individual games or aggregated across multiple games or days to track overall progress.
This was all done over a 12-month period.
I trained using sorted texts of two alignments: Scum and Town.
Forum software (knowledge of this needed for efficient scraping)
I've been doing a bit of research as of late. To be more specific, some (relatively) intense web scraping has been performed on this site, WG, Mafia Universe, and The Syndicate, and especially MafiaScum, over the past year or so in tandem with this dataset (which contains many games obtained from here). I will not be divulging much, depending on what one would define as much (no graphs, not including complete methodology and key data, so a lot of this could be considered by some to be a "whole lot of nothing"). Obviously (as noticed from the past few games) my usage of this data and understanding of my tells hasn't truly been exactly the best. I will formally be putting my training/testing to the test in a game I'm participating in soon, as I believe I have achieved sufficiently good accuracy.
Some things I've done
- Qualitative finding (personal): Analyzed common scum tells and looked at the reasons behind why they their kills in order to try to figure out ways to play while minimizing their chances of getting nightkilled. Newcomb's tell only bears the desired degree of accuracy when the thought process behind it is also considered, from my experience.
- I examined some games from 2018 - 2020 from MafiaScum (randomly chosen range). I wanted to determine posting frequency of scum vs. town. for this isolated data set, I only used posts on D1, so I didn't have to adjust for shifting Town:Scum ratios on future days. On D1, scum make up a shade over 23% of the player pool, so we should expect each post to have a ~23% chance of being from scum. Across 48,094 D1 posts, 8,921 were from scum (18.54%) and 39,173 were from town (81.45%).
A Few of My Findings
- Wolves tend to achieve victory (~73% chance from my Syndicate dataset) when within the top 25% percentile of posters.
- If two people are suspecting each other for arbitrary reasons at the start of the game, there is a statistically significant chance above the norm that they're partners.
- Certain game activity patterns are alignment-indicative (not specifying).
- There are major differences in deceptive behavior between people of different genders (gender automatically obtained from Mafia Universe).
- Scum tended to use language that was simpler in structure, using shorter sentences and bearing fewer details, making it easier to read. However, this simplification did not extend to their vocabulary variety, as their lexical diversity was found to be similar to that of townies.
- Scum wrote longer posts and wrote shorter sentences.
- Scum may self-vote when being bussed by fellow teammates.
------
Some of the findings I realized (vast majority obviously not seen here) are built upon the findings of retired mafia player Ellibereth.
None of this is to say my play these past few months has been all of a greater scheme to refine and find better tells to obtain better reads and all, or that I have been trolling in focus of research and not playing to the best of my ability; it is ultimately the result of subpar play, tunneling, and pushes. These last few games only built upon me testing the efficacy and usefulness of common tells on a surface-level.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
AI
As to why I've never engaged with, and stonewalled the AI accusations:1) I believe it decreased hostility as opposed to what it would be if I had responded.
2) In order to add to my pool of data to assess the aggressiveness as an effective scum-tell when applied on the surface (thought process not considered).
3) To read Lethal (especially Lethal, as I scraped many of their posts and noticed their past AI accusations against others (likely not serious), The X, and others with a certain degree of accuracy.
Examining read accuracy (using same method as Ellibereth)
To measure the accuracy of my reads and track improvement, I used the Brier Skill Score (BSS). This method compares the accuracy of my probabilistic reads (e.g., stating Player A is 80% Town) against a baseline "no skill" reference forecast. This "no skill" forecast represents someone guessing alignments based only on the known number of players for each alignment at that point in the game (e.g., if 4 players remain with 3 Town and 1 Mafia, the no-skill guess is 75% Town for everyone).
The calculation involves first finding the Brier Score (BS) for my set of reads. The Brier Score is the mean squared difference between my forecasted probability (ft) and the actual outcome (ot, where 1 means the predicted alignment was correct, and 0 means it was incorrect). The formula is:
BS = (1/N) * Σ(ft - ot)²
Here, N is the total number of reads I'm evaluating. A lower Brier Score is better, with 0 being a perfect score.
Next, I calculated the Brier Score for the "no skill" reference (BSref) over the same set of players/reads, using the base rate probabilities (derived from player counts per alignment) as the forecast (ft) for each read in the formula.
Finally, I computed my Brier Skill Score (BSS) using my Brier Score (BSf) and the reference Brier Score (BSref):
BSS = 1 - (BSf / BSref)
A BSS of 1 means my reads were perfect. A BSS of 0 means my reads were exactly as accurate as the "no skill" guesser. A positive BSS (between 0 and 1) indicates my reads were better than the baseline, while a negative BSS means my reads were worse than simply guessing based on the numbers. While this method works best for probabilistic reads, it can be adapted for binary reads ("Player is Town" or "Player is Mafia") by treating them as 100% certainty (1.0 or 0.0 probability), though this often results in poorer scores compared to nuanced probabilistic reads. This calculation can be done for individual games or aggregated across multiple games or days to track overall progress.
This was all done over a 12-month period.
Not mentioning achieved BSS score. You can find out.
I trained using sorted texts of two alignments: Scum and Town.
Forum software (knowledge of this needed for efficient scraping)
- WorstGen: XenForo
- Mafia Universe: vBulletin
- The Syndicate: phpBB
- MafiaScum: phpBB
Last edited:
