On 29 June, the MyTracker team took part in FinChess, a chess tournament among financial services and IT companies. As part of FinChess Public talks, Alexander Smirnov, Research Programmer of the MyTracker predictive analytics team, spoke about the similarity of chess algorithms and machine learning models used in MyTracker.
This article presents the key takeaways of his report, divided into three sections:
We will be looking at some exciting parallels between a game of chess and an IT product, and into how MyTracker's Personalize, LTV Prediction and Fraud Scanner solutions work. It will be useful for analysts, marketers and chess fans.
An increase in computational powers was a game changer for both chess and today's IT, helping to put into action numerical modeling techniques accumulated for decades.
This gave rise to chess engines that learned to cope with resource-intensive tasks such as recommendations for the best move.
A recommendation for the next move is a calculation and a selection of a combination of moves most likely to lead to a desirable outcome – victory in the game.
This is also a job of any recommendation service, for instance:
Recommendations by chess engines
Recommend moves most likely to lead to victory in the game.
Select goods most likely to lead to a purchase or other desirable action.
The number of moves is finite, which is why the system can go through all of them and find ones that will lead to the desired outcome. As this work is ultimately about computing powers and an exponential increase in the number of moves to be calculated, advanced chess engines as well as recommendation services rely on reinforcement learning.
Reinforcement learning is a technique used to train algorithms to interact with an environment and learn from the feedback. If an action leads to a reward, algorithms will perform that action or sequence of actions much more often.
Examples of rewards:
This principle underpins AlphaGo Zero, a learning algorithm that delivered surprising results. In 2017, it beat Stockfish, one of the strongest traditional engines. AlphaGo Zero played a huge number of games against itself and learned to select the best moves in most situations.
Multi-armed bandit algorithms, often used in recommendation services (such as MyTracker Personalize) to select the best product (offer), are also built on reinforcement learning.
After making an offer, the bandit receives feedback in the form of a payment made or not made by the user. This way, the bandit gets to know how its actions (offer displays) influence the environment (user payment behavior) and comes up with the best offer.
To learn more about the MyTracker Personalize recommendation service, read our case on the Hustle Castle mobile game.
Before the game, the chess player and their team analyze the opponent's actions (moves). At this stage, they collect information about the opponent – their playing style, recent games, preferred moves and weaknesses.
Predictive analytics has a similar objective when analyzing the ROI of an advertising campaign. The sooner you get this information, the earlier you will be able to end an unsuccessful campaign and save your money. Just like in chess, data is obtained by reviewing previous advertising statistics – audience covered, progress, and earnings.
Prediction in chess
Analyzing the opponent's previous games to learn about their playing style and make assumptions about how they will play the next game
Analyzing data on earnings and progress of the ad campaign to make assumptions on future ROI
Models built on historical data may be instrumental in making faster and more accurate predictions. For ad campaigns, one can use simple linear models, while for chess, maps of probable moves will serve the purpose.
To learn more about how to build your own LTV prediction for a mobile app, see our guide.
To make sure your predictive model is as effective as possible, you may need an additional expert review that will factor in aspects overlooked in the model.
For instance, a chess expert may note that the analyzed games were played with weak opponents, hence the moves will not repeat in a real game. For an ad campaign, an expert may identify internal processes that had an impact on the campaign, such as seasonal activities or launch of new functions.
An expert review is always available to MyTracker customers to assess predictive analytical data. To use the service, contact your manager or customer support.
The key challenge for predictive analytics and preparations for a chess tournament is a sudden change of circumstances:
In this case, current conditions no longer conform to historical data used as inputs for training. As a result, predictive analytics loses its accuracy and value. At some point, however, these situations will also become part of historical data to which algorithms will adapt.
This was the case with iOS 14.5+ updates, which initially reduced the value of predictive analytics to zero. However, later on, the services learned to make confident LTV predictions even with limited inputs.
Learn more about how to predict LTV in iOS 14.5+ on our blog.
Foul play exists both in chess and IT. In chess, it is called cheating, while in IT, it is fraud. General approaches to defining foul play in these two areas are very similar.
As chess engines progressively gained strength, special tournaments were organized for them to compete in. They did so well that in 1997, Deep Blue beat the then world chess champion Garry Kasparov. Over the course of time, their advantage over humans kept growing.
Then a chess engine, which remained a driving force in chess, became a cheating tool. This problem was especially relevant in online matches.
In human chess matches, use of engines is prohibited. Players can employ them as a strong opponent in training, but not in an actual game.
If one player uses a chess engine, they will have a clear advantage and the match will be unfair. If both players use chess engines, this is a competition of algorithms, not humans.
The purpose of in-app advertising is to attract users that will install the app and perform certain actions – purchase, subscribe or just make daily visits and watch ads for a month.
Ad partners – services placing ads on online platforms (websites, video hosting sites and other apps) – attract users to the app, and that's where a fraud may occur. Ad partners may cheat to get money for new users that turn out to be bots.
Bots that visit the app instead of real users may perform the required target actions (placing an order, reaching a certain level in the game), sufficient for the partner to reap a reward. After that, bots stop their activity and bring no profit to the app.
The behavior of cheaters and fraudsters differs from that of ordinary players and users.
Cheater in chess
Fraudster in advertising
A cheater uses a chess engine to outplay their opponents – make quicker decisions, predict their rival's every move and make practically no mistakes.
Bots are either too quick or too slow to press the install button, which betrays their abnormal behavior within the app.
Fraudsters try hard to disguise themselves, and anti-fraud systems employ increasingly sophisticated solutions to distinguish between a good and a bad user.
At any tournament, all moves are recorded. Modern chess engines make a list of the most favorable moves for each position in the game. If a human's moves fully replicate those advised by the computer, this may be cheating.
Chess engines can analyze the accuracy of the player's move, that is, how closely that move followed the engine's recommendation. After the game, the accuracy of moves is often measured, which can help detect a cheater.
Many online chess platforms calculate this parameter automatically following each game.
If the person is seen to play with low (or average) accuracy in previous games and then, within a short time frame, they suddenly win several games in a row with an abnormally high accuracy, they may be cheating.
Another sign of cheating is when a player completed the last thousand games with a 40–50% accuracy, but during the last dozen games their accuracy score went up to 70%. A smart cheater may ignore the strongest moves, using number 3 and 4 computer recommendations instead. This may still be enough to win, but cheating will be harder to detect.
In advertising, anti-fraud systems, such as Fraud Scanner from MyTracker, serve to detect fraud. They do the following:
To illustrate this, let's look at the Time to Install parameter, which is the time between going from the ad to the page in the app store and pressing the install button.
The upper chart shows normal time distribution before install obtained from the traffic of a reliable partner. We will use it as a reference. On the lower chart, there is an abnormal peak, indicating that a certain percentage of users press the install button too quickly. This is a sign of fake traffic.
When detecting abnormalities, the anti-fraud system warns about potential fraud. These data can be further checked by means of additional reports and used in disputes with the ad partner to return money for bot clicks.
In this article, we outlined three explicit links between chess and MyTracker digital services:
This knowledge will give you a better insight into how IT solutions powered by machine learning work and how you can use them more effectively to develop your own products.
To start working with us, simply enable MyTracker.