How artificial intelligence technologies help Aviasales grow: seven examples
At the conference of Product Marketing of Epic Growth Conference Head A of Mobile Hotels is in Aviasales Konstantin Savchenko told about the experiments and examples of the use of artificial intelligence technology in Aviasales.
Watch the video or read the notes under the cut.
Many of the solutions in Aviasales that are related to artificial intelligence technologies have grown out of hackathons. And the majority of these solutions from the very first version give tangible growth and good business results, for example, an increase in conversion or a reduction in costs.
To use the technology of artificial intelligence, it is not necessary to be in this pro. The secret here is simple: large corporations have already done everything for you. And laid out, for example, ready-made solutions on GitHub, where you can find neural networks and smart libraries.
To inspire you to try creating solutions based on artificial intelligence, Konstantin Savchenko gathered seven examples of how these technologies are used in Aviasales.
The order of the partners in the ticket is a simple example, but once it helped a lot to start to understand what machine learning is. This is the ticket that you find on Aviasales. For a specific ticket different partners provide their prices.
Often prices from different partners are the same. We need to choose which partners to put on the big orange “Buy” button, which most users click. Of course, first of all we put the lowest price, the only way it works. But if there are several low prices from different partners and they are the same, we need to choose the best one.
In this case, we focus on two parameters. The first is the conversion from clicking the buy button into the purchase. And the second is the commission that a particular partner pays us. For each, a label is drawn up (see the screen below), which helps in the first place to determine the partner with maximum efficiency.
All our partners want to improve their funnels, so they do a lot of experiments, and the conversion periodically changes. It is important to keep track of this and this is the moment that can be automated.
Suppose that in 5% of cases you bet on the “buy” button of a non-partner with the best price, you start chasing all other partners to find out what their conversion is now. You update this label, recalculate the productivity - and, thus, the next user already sees the new order of partners. Your system is trained on the data it receives from partners, and selects the best solution. This can be called machine learning.
# 2 Sort hotels
If everything is simple enough with tickets: you can sort them by price and put the cheapest one in the beginning, then this reception will not work with hotels.
If we show the cheapest hotel, most likely it will be a hostel on the outskirts and hardly anyone will like it. You can start doing the same thing as with the tickets: show all the hotels one by one, look at their conversion and choose the best ones. But we have 4 million hotels. I am afraid that none of us will live to see the results of this test. Therefore, we resort to using artificial intelligence technologies.
There is also a turnkey solution. In this case, the “smart” library, which was made by the guys from “Yandex”, was created just for those who still do not understand artificial intelligence. Hotels have a large number of characteristics on the basis of which the user makes his choice: price, rating, reviews, and so on. At the entrance you give the library the parameters of the hotel; It turns out to transfer conversions from showing the hotel to the purchase.
What does this library do? She tries on the basis of these data to predict what conversion will be in similar hotels. At the output, you get a conversion forecast that can be used as a sort.
In this experiment, we have increased the average check + 17%. This algorithm began to show more expensive hotels above the rest - and thus people began to buy more expensive hotels.
Other indicators and everything related to conversion increased significantly: conversion to sales + 6%, income + 19%.
# 3 Photo Analysis
Partners provide us with many photos for each hotel. But we do not know what is depicted on them. We need AI to know what quality they are and in what order they need to be shown. Among the photos there are such:
This famous hair dryer somehow got into the top issue in Moscow. This was one of the reasons why we decided to figure it out.
There are a huge number of libraries; we found a suitable one that is trying to determine the location shown in the photo.
We drove all our photos through this library (you can call it a trained neural network) and got the result - an approximate breakdown of what the library sees in the photo.
It was important for us to understand whether it is on the street or inside. If on the street, we were primarily interested in the pool. Inside - beds, toilets, hall.
Then we decided that users are interested in seeing first of all the hotel room. What is a number? This is when the picture shows a large bed. To cope with this was not very difficult. We started to manually review what happened: everything looked cool, but at the resort areas (especially for mass tourism) the photos of the beds looked bad. It was a very poor bed in a very poor room.
We analyzed what our partners and competitors are doing in this case. And they show photos of the pool, because the pool in these hotels is always beautiful. We began to put forward precisely those hotels that have a beautiful photo of the pool.
By running this issue, we not only got rid of manual labor (we used to hire people on freelance, who selected photos of hotels in top cities), but also increased the conversion by + 12%, which increased mainly due to beach locations in the pool experiment.
# 4 Review Analysis
The aesthetics of the photographs and the style of the interior are what we can work with, as we thought. Often very similar in characteristics hotels are made in a completely different style. You can find out where the interior is - not only by photos, but also by reviews.
Users often write about how they like the interior. I didn’t meet a lot of reviews, for example: “This is an amazing room like my grandmother’s”. But about some modern and stylish hotels, users usually write. They write about the location, proximity to attractions or the view from the window.
When users search for a hotel, they first sift out everything that does not suit them, leaving several options as favorites. And the next step that influences the choice is viewing reviews. There are often too many reviews. We thought it would be cool to read the squeeze, that is the most important thing. From this idea and started.
Attracted our partners who specialize in the analysis of reviews. Together with them we pulled the most important things from the reviews and collected a certain set of badges that we placed on the hotels.
We really wanted to launch this feature, we dreamed very much about it. But it turned out that people do not care. We hung beautiful badges on hotels that revealed the main advantage of the hotel. But this had no effect on conversion and on numbers.
# 5 Predicting Ticket Prices
For all the time in Aviasales, we have accumulated a huge amount of data. And our hypothesis was that there is a relationship between how the price of tickets varies depending on how much time is left before departure or on what day this departure.
This was also one of our hackathon projects, where the guys developed a solution that quickly began to give cool results.
Thanks to this solution, we began to save on data, began to fill in the price calendar those places, prices and dates for which we had no real data.
It works with incredible accuracy: it is only mistaken for 10% of the price, which seems to be a good indicator for a decision made on the knee.
What else is interesting with predictions? People often decide whether to wait for a drop in ticket prices or still worth buying now. Thus, we began to make prompts for users to "buy now" or "wait." Usually ticket prices only grow, so in 90% of cases we say: “Buy now”. User confidence here was minimal.
Below is a layout of what we plan to do. We will show the graphs of how the price will change according to our forecasts. We expect to get more user confidence from this.
# 6 Predicting Hotel Cancellations
Most users buy non-refundable tickets and the fact that the user bought a ticket can be considered the final deal.
In the case of hotels in a different way; The share of returns is high and it is important for us to plan how much money we will earn here. Therefore, on the basis of how much time is left before booking and on the basis of past user actions, we try to predict what the percentage and price of cancellation will be. It helps in planning.
# 7 Traffic Quality Assessment
Most often, people travel twice a year. Therefore, when they install the application, it is not at all a fact (and this is normal) that they will not buy tickets now. But it is still important to assess how good this source of traffic is. We are trying on the first actions of the user to predict what the probability that he will make a purchase.
I want to draw your attention to the first three points. Thanks to these points, it seems to me that we have learned that it is quite simple to implement artificial intelligence technologies. I recommend taking your developer and spending one day on research.
If you have any task that you think you could automate. There is a great chance that what you need to do has already been done before you, it will not take long to apply it to yourself.
More product marketing reports available at the @epicgrowth Telegram Channel .
Decoding performances published on vc.ru.
Watch the video or read the notes under the cut.
Many of the solutions in Aviasales that are related to artificial intelligence technologies have grown out of hackathons. And the majority of these solutions from the very first version give tangible growth and good business results, for example, an increase in conversion or a reduction in costs.
To use the technology of artificial intelligence, it is not necessary to be in this pro. The secret here is simple: large corporations have already done everything for you. And laid out, for example, ready-made solutions on GitHub, where you can find neural networks and smart libraries.
To inspire you to try creating solutions based on artificial intelligence, Konstantin Savchenko gathered seven examples of how these technologies are used in Aviasales.
# 1 Order of partners in the ticket
The order of the partners in the ticket is a simple example, but once it helped a lot to start to understand what machine learning is. This is the ticket that you find on Aviasales. For a specific ticket different partners provide their prices.
Often prices from different partners are the same. We need to choose which partners to put on the big orange “Buy” button, which most users click. Of course, first of all we put the lowest price, the only way it works. But if there are several low prices from different partners and they are the same, we need to choose the best one.
In this case, we focus on two parameters. The first is the conversion from clicking the buy button into the purchase. And the second is the commission that a particular partner pays us. For each, a label is drawn up (see the screen below), which helps in the first place to determine the partner with maximum efficiency.
All our partners want to improve their funnels, so they do a lot of experiments, and the conversion periodically changes. It is important to keep track of this and this is the moment that can be automated.
Suppose that in 5% of cases you bet on the “buy” button of a non-partner with the best price, you start chasing all other partners to find out what their conversion is now. You update this label, recalculate the productivity - and, thus, the next user already sees the new order of partners. Your system is trained on the data it receives from partners, and selects the best solution. This can be called machine learning.
# 2 Sort hotels
If everything is simple enough with tickets: you can sort them by price and put the cheapest one in the beginning, then this reception will not work with hotels.
If we show the cheapest hotel, most likely it will be a hostel on the outskirts and hardly anyone will like it. You can start doing the same thing as with the tickets: show all the hotels one by one, look at their conversion and choose the best ones. But we have 4 million hotels. I am afraid that none of us will live to see the results of this test. Therefore, we resort to using artificial intelligence technologies.
There is also a turnkey solution. In this case, the “smart” library, which was made by the guys from “Yandex”, was created just for those who still do not understand artificial intelligence. Hotels have a large number of characteristics on the basis of which the user makes his choice: price, rating, reviews, and so on. At the entrance you give the library the parameters of the hotel; It turns out to transfer conversions from showing the hotel to the purchase.
What does this library do? She tries on the basis of these data to predict what conversion will be in similar hotels. At the output, you get a conversion forecast that can be used as a sort.
In this experiment, we have increased the average check + 17%. This algorithm began to show more expensive hotels above the rest - and thus people began to buy more expensive hotels.
Other indicators and everything related to conversion increased significantly: conversion to sales + 6%, income + 19%.
# 3 Photo Analysis
Partners provide us with many photos for each hotel. But we do not know what is depicted on them. We need AI to know what quality they are and in what order they need to be shown. Among the photos there are such:
This famous hair dryer somehow got into the top issue in Moscow. This was one of the reasons why we decided to figure it out.
There are a huge number of libraries; we found a suitable one that is trying to determine the location shown in the photo.
We drove all our photos through this library (you can call it a trained neural network) and got the result - an approximate breakdown of what the library sees in the photo.
It was important for us to understand whether it is on the street or inside. If on the street, we were primarily interested in the pool. Inside - beds, toilets, hall.
Then we decided that users are interested in seeing first of all the hotel room. What is a number? This is when the picture shows a large bed. To cope with this was not very difficult. We started to manually review what happened: everything looked cool, but at the resort areas (especially for mass tourism) the photos of the beds looked bad. It was a very poor bed in a very poor room.
We analyzed what our partners and competitors are doing in this case. And they show photos of the pool, because the pool in these hotels is always beautiful. We began to put forward precisely those hotels that have a beautiful photo of the pool.
By running this issue, we not only got rid of manual labor (we used to hire people on freelance, who selected photos of hotels in top cities), but also increased the conversion by + 12%, which increased mainly due to beach locations in the pool experiment.
# 4 Review Analysis
The aesthetics of the photographs and the style of the interior are what we can work with, as we thought. Often very similar in characteristics hotels are made in a completely different style. You can find out where the interior is - not only by photos, but also by reviews.
Users often write about how they like the interior. I didn’t meet a lot of reviews, for example: “This is an amazing room like my grandmother’s”. But about some modern and stylish hotels, users usually write. They write about the location, proximity to attractions or the view from the window.
When users search for a hotel, they first sift out everything that does not suit them, leaving several options as favorites. And the next step that influences the choice is viewing reviews. There are often too many reviews. We thought it would be cool to read the squeeze, that is the most important thing. From this idea and started.
Attracted our partners who specialize in the analysis of reviews. Together with them we pulled the most important things from the reviews and collected a certain set of badges that we placed on the hotels.
We really wanted to launch this feature, we dreamed very much about it. But it turned out that people do not care. We hung beautiful badges on hotels that revealed the main advantage of the hotel. But this had no effect on conversion and on numbers.
# 5 Predicting Ticket Prices
For all the time in Aviasales, we have accumulated a huge amount of data. And our hypothesis was that there is a relationship between how the price of tickets varies depending on how much time is left before departure or on what day this departure.
This was also one of our hackathon projects, where the guys developed a solution that quickly began to give cool results.
Thanks to this solution, we began to save on data, began to fill in the price calendar those places, prices and dates for which we had no real data.
It works with incredible accuracy: it is only mistaken for 10% of the price, which seems to be a good indicator for a decision made on the knee.
What else is interesting with predictions? People often decide whether to wait for a drop in ticket prices or still worth buying now. Thus, we began to make prompts for users to "buy now" or "wait." Usually ticket prices only grow, so in 90% of cases we say: “Buy now”. User confidence here was minimal.
Below is a layout of what we plan to do. We will show the graphs of how the price will change according to our forecasts. We expect to get more user confidence from this.
# 6 Predicting Hotel Cancellations
Most users buy non-refundable tickets and the fact that the user bought a ticket can be considered the final deal.
In the case of hotels in a different way; The share of returns is high and it is important for us to plan how much money we will earn here. Therefore, on the basis of how much time is left before booking and on the basis of past user actions, we try to predict what the percentage and price of cancellation will be. It helps in planning.
# 7 Traffic Quality Assessment
Most often, people travel twice a year. Therefore, when they install the application, it is not at all a fact (and this is normal) that they will not buy tickets now. But it is still important to assess how good this source of traffic is. We are trying on the first actions of the user to predict what the probability that he will make a purchase.
Seven examples
- sorting partners;
- sorting hotels;
- photo analysis;
- review reviews;
- price prediction;
- prediction of cancellations;
- traffic estimate.
I want to draw your attention to the first three points. Thanks to these points, it seems to me that we have learned that it is quite simple to implement artificial intelligence technologies. I recommend taking your developer and spending one day on research.
If you have any task that you think you could automate. There is a great chance that what you need to do has already been done before you, it will not take long to apply it to yourself.
More product marketing reports available at the @epicgrowth Telegram Channel .
Decoding performances published on vc.ru.