Solo In-Lane Match-Ups

Role Contribution to Class Classes Derived From Match-Ups

Ok real talk. I didn’t ONLY choose to start with Support and Adcs because I’m a Support Main. I started there because I observed (by rolling around i the data) that those classes are the most stable across all positional interactions. However, I want to take a moment to quantify this observation and offer some potential narratives before we look at the Top and Mid in-lane classes.

1. How much does each role interact with each other role?

This plot shows how many percentage points of win rate for one role is explained by a match-up in the opposing role. More technically, it’s the median of the absolute value of all of roleX’s match-ups with roleY. Basically, if the number is 0.007 at the intersect of Sup and Jun, that means that the median match-up between a Support and Jungler is a 0.7% increase or decrease in WR.

A few things jump out at us.

• Jungle has equivalently strong match-ups across all roles. Unsurprising, as jungle interacts with everyone.
• Adc and Sup also have similar match-up strength across all roles. This makes sense as Supports roam and Adcs exert most of their agency in the late game where they must interact with everyone.
• Top and Mid have much stronger match-ups swings vs their lane opponents, especially top. Top has the highest ratio of in-lane match-up strength to out-of-lane match-up strength of any role. Interpretation: Top and Mid need to win lane and struggle to exert agency when they do not.

2. How much does each role’s match-ups with each other role contribute to the UMAP clusters, and therefor derived classes?

This is a hard question to answer and I approached it in a messy way, but I think it gets the job done. In these blogs we’ve been using champion match-ups to inform UMAPs that cluster champions with similar mach-ups together (we used same approach with synergies). In the Bot lane, I used the in-lane match-up (Sup vs Bot) for one clustering and the complete suite of match-ups (Sup vs all roles) for a second clustering. But how much does each other role’s match-ups contribute to the overall clustering? How much is Sup clustering affected by Top for example?

To get at this, I made UMAPs using single sets of match-ups (Top vs Top, Top vs Jun, Top vs Mid,… etc) and determined how well they correlated with UMAPs made from all match-ups (Top vs. All). Higher values mean that the UMAP made with only that role’s match-ups is more similar to the UMAP made with all match-ups (1 is matching, 0 is completely uncorrelated). TLDR: Higher values mean that match-ups with that role contribute more to the overall UMAP and classes.

We see some interesting phenomena:

• Support clusters (and by extension the derived classes) look almost the same no matter which role’s match-ups we use to make them. We saw this in our first Bot Lane Ecosystem post when the UMAP using Bot match-ups near-perfectly matched the UMAP using all match-ups.
• Conversely, Top vs Top match-ups near-perfectly matches the Top vs. all UMAP. Conversely, Top vs Sup/Adc/Mid/Jun all are quite uncorrelated with Top vs all. At the same time, Sup and Adcs can recapitulate their vs all UMAPs well using vs Top only. Interpretation: Top laners have completely different classes when they are on their island than when they are off it.
• Mid-lane trends are like Top-lane-light.
• Jungle is more affected by its match-up with Mid than with Jun for determining Jun classes.

I wanted to dig into this further, so I also tried leaving out a role’s match-ups and seeing how well the UMAP correlates with the UMAP using all match-up data. While the last plot used just one set of matchups, this leaves just one set (the one on the X axis) out. Higher values mean leaving that role out mattered less.

• Top lane classes are DOMINATED by Top vs Top interaction. This makes a ton of sense given how Top Match-ups are more than double the magnitude of non-Top match-ups (as shown in heatmap 1).
• Mid lane is again Top-lane-light.
• Sup match-ups are more important for defining Adc classes than any other role.
• Jun and Sup generate similar UMAPs when any 1 role match-ups are not included.

Because Top and Mid global match-ups are so different from their lane match-ups, i am going to cover their global match-ups in a later post where we can compare all roles global match-ups at once. With that out of the way, let’s get into the solo lane role match-ups!

Disclaimer: UMAPs are stochastic in nature so I did bootstrap my UMAP correlations and used the mean, but I recognize this method is imperfect and I’m happy to try others if someone has a more elegant suggestion. FYI the Pearson is of each champion’s distance to each other champion on the UMAP compared between the two relevant conditions. An identical UMAP would form a line. I am aware of the limited relevance of distances between distant members of a UMAP and that certainly limits the value of this analysis, but I think it is fine for extracting general trends as we did here.

Top vs Top Match-Ups

There are more champions and potentially more complicated relationships between the clusters in Top, so I added a third dimension to the UMAP to give it some more dimensional space to work with. However it still works the same, closer together means more similar lane match-ups.

Psst… You can zoom in and out of the figure, rotate it, and turn classes on and off by clicking on them in the legend!

Just in case the 3D version doesn’t work well, here’s 2D version:

We see some interesting groups forming! Let’s see how they interact:

• Drain Fighter: With enough sustain to withstand most tank match-ups and some ranged match-ups, this class is weak to fighters who usually have more damage.
• Fighter: Dominates Drain Fighters and reasonable vs. any non-ranged match-up, their weakness is their short threat range.
• Neutralizers: Natural enemy of ranged champions, each one has a way to ignore or kill most ranged champions. This burst and mobility comes at the cost of sustained tank killing, and thus vs. tanks they are weak.
• Ranged: They prey on the immobility of fighters, struggle to fully poke out tanks, and are susceptible to the mobility and burst neutralizers have.
• Tanks: They crush neutralizers and are mixed vs other classes that don’t have the sustain of Drain Fighters.

I was surprised to see Top divide into so few classes, but the interaction matrix looks pretty clean and strong. I think the Fighter class could have been further divided, but I think these classes seem intuitive and will hopefully be a helpful and memorable shorthand in champ select.

Mid Lane Match-Ups

Now let’s take a look at Mid Lane in-lane match-ups the same way.

And again, in case the 3D version doesn’t work, here is a 2D version:

Mid separates into some very distinct and recognizable clusters. That said, after looking at their interactions I did expand the number of classes to 8. Near the top we see the mages separated by their range and utility. On the bottom we see AD and AP assassins.

Let’s take a look at the class interactions

• Mid ADCs: They are good vs AP assassins by abusing their melee ranges early. They are wrecked by the long range of Artillery Mages.
• AD Assassins and AP Assassin: They are best vs. mid and long-range mages. They are weak to Counter-Assassins.
• Ranged Assassins: They are catchers who have generally neutral match-ups, though they are a bit weaker vs. bot AD and AP Assassins.
• Counter-Assassins: They shut down the dash- and burst-dependent assassins, but they get hosed by the range and sustained damage of mages.
• Defensive Mages: They have the range and DPS to deal with ADCs, the defensiveness to somewhat resist assassins, but they are weak to the longer-range mages who can poke them out of lane.
• Mid-Range Mages: Weak to assassins due to lack of defense, strong vs Counter-Assassins and Defensive Mages due to their poke. They are also weak to Artillery Mages who do what they do but longer-range.
• Artillery Mages: They crush everything that doesn’t get in their personal space to murder them.

We again see clean, well-behaving classes. Still, there are a few unlikely bedfellows. Tristana and Akshan behave more like AD Assassins than ADCs, so I reclassified them. Leblanc performs more like an AP Assassin (most of whom are melee) than a Ranged Assassin. Most others make perfect sense though and follow trends we would intuit.

You can also observe what I like to call “class-killers”. If you look along the diagonal you can identify champions that excel against their own class. For example, Xerath is the king of Artillery Mages, Yone excels vs. most other AD Assassins, and Kassadin is king of the AP Assassins.

The Future

Outstanding on my to-do list is to look at Jungle match-ups, but because Jungle is so dependent on interactions on other roles to form classes, it makes sense to do one mega-post for Jungle and all role global match-ups and classes. Some roles like Support and Adc will see their roles largely unchanged, while some roles like Top will see wild reshuffling. I also need to do a similar post for synergies.

As a shorter-term goal, I will be releasing a “Bot Lane Expanded Universe” post soon to class the low-playrate champions like Mage ADCs.

Anyway that’s all for now. Let me know if the 3D UMAPs are too cumbersome. I may revert to 2D in future. 3D does give more space for clusters to interact at their fringes, but it comes at the cost of readability.