How to start using artificial intelligence
It seems that AI is very difficult, and in a small company
But it is not always necessary to compose something difficult.on their own. Big companies have already invented everything and put it in open access on GitHub. Both neural networks and smart libraries can be found there. For developers, this is a great opportunity to try new things and peek at how others solved the problem.
At Aviasales, many solutions related to AI technologies are invented during internal hackathons.
Hackathonis a competition among developers whenit is necessary to solve some problem in a very short time, for example, 48 hours. Naturally, during this time it is impossible to create something from scratch, so ready-made solutions are used.
The most important thing - fast experiments with ready-made technologies almost always show good results, be it an increase in conversion or a reduction in costs.
"The Prophet" predicts when to buy a ticket
During one of the hackathons, a hypothesis appeared thatthere is a connection between the time before departure and the day you start the trip. Having analyzed a huge numberstructured data that has accumulated over 11 years of operation of the Aviasales service, it was possible to prove that the hypothesis is correct. This is how the Prophet service appeared, which predicts the best moment to buy tickets with an error of 10%.
Thanks to the new service, the company began to saveby obtaining third-party data and was able to insert prices into the calendar for those dates and destinations where there was no real data - with a small amount of error, the Prophet helps to find out the price in advance.
The “Prophet” gives travelers tips abouttime to search for tickets: “Buy now” or “Wait.” Along with the search words, a graph is shown showing how the price will behave based on the company’s forecasts.
AI selects the best ticket seller
In the Aviasales meta-search, 200 ticket offices are presented.and 728 airlines. It is clear that in the first place is always the ticket with the lowest price. But a ticket can have several sellers, and often some have the same value. Then the question arises: who should be ahead?
The yellow button "Buy" - this is the first place among all sellers. Under the button is a list of agencies andairlines where you can also buy this ticket: for the same price or more expensive. To determine who to put on the magic button, two factors are taken into account - the commission that the partner pays for the ticket sold, and the conversion from going to the seller’s site to the purchase. That is, these are factors that take into account the interests of two parties - the meta-search and the convenience of the traveler.
All data on both factors are recorded inthe table. The data is constantly changing, as sellers are working to improve their sites. It was decided to automate this process in order not to enter the numbers into the table manually. So, in 5% of cases, on the “Buy” button, the seller turns out to be with not the lowest price to find out what percentage of users will go to his website and buy a ticket. Thus, the parameters are recalculated all the time, the system is trained on the basis of the obtained data and chooses the best solution itself.
AI chooses photos for the hotel description
If the choice of product or service is related to qualityphotographs, and there are a huge number of them, it is irrational to select pictures manually. Need AI. The problem is that each partner sends their photos of the hotel to the Aviasales hotel service, and partners are not always chain giants like Hilton or Marriott. Sometimes this is the owner of a small guesthouse in Crimea, who photographed the rooms on his phone.
To analyze photos you need AI, whichrecognizes quality and determines in what order to display images. The solution was found in one trained neural network that can determine location. The result is, for example, the following breakdown: 63% - building, 20% - pool, 11% - tree, 6% - beach.
In city hotels it is interesting how the room lookstherefore, photos from the bed are shown first. In beach hotels, on the contrary, the pool and sun beds are important. As a rule, in resort areas the numbers are rather scarce, and the interior of the room is best shown last.
Starting to work with photos using AI,the company has reduced the cost of manual labor: previously hired freelancers who took pictures in popular cities and also increased conversion by 12%, mainly due to experiments with photographs of swimming pools at beach resorts.
How AI helps make beautiful websites with Weblium designer
Weblium website builder uses AIDesign Supervisor, which tracks the user's website creation activities in real time and identifies design errors, correcting them on the fly.
Because the entire project uses a product stackGoogle, then the developers used Google Cloud AI to implement this task. The most difficult task was to teach the neural network to see design problems with incorrect use of colors, font pairs, and the like.
To obtain a large enough data set,the developers trained the model on a data set with 30 million design solutions taken from the leading resources Behance and Dribbble. The structures of sites and elements were recognized using the Cloud Vision API. This allowed us to make a “quantum leap” in achieving the accuracy of AI Design Supervisor.
We can not yet boast that AI DesignSupervisor works flawlessly, but it can already be accurately used as the main point of differentiation from competitors. Users constantly write that turning one site into another on the fly invariably causes a wow effect even when AI Design Supervisor is used repeatedly.
David Brown, founder of Weblium
At the same time, Weblium is working tocontextually determine the content that the user contributes to sites, understand his tasks and offer him the most relevant blocks when building the site. To do this, developers use the Cloud Natural Language API.
And the latest development, very important inperspective - voice interfaces. Weblium AI Lab prototypes the voice control of the site builder using the Cloud Speech-to-Text library. The final idea is that the user could put a technical task in a voice and in rather simple words, for example: “I want a modern functional site for my car wash”. And as a result of this TK, you get a decent site.
How Sephora and Lamoda AIs are used
Dynamic Yield's personalization platform helps partners improve the customer experience. It is used by such famous brands as Sephora and Lamoda.
Dynamic Yield can segment your audience,select personalized products and content. The platform works on the web, on mobile devices, and can be used to send newsletters and place advertisements. It delivers personalized recommendations to users across all communication channels.
Sephora tested personalization systemrecommendations in eight online stores in Asia. In each of them, recommended products were selected for users, guided by three strategies: similar products, related products, and automatic recommendations.
Until the introduction of AI, the final choiceThe products that will be shown to the user were made depending on the country and KPI. Now they are shown depending on what products the user added to the cart and which ones he ultimately purchased.
Thanks to this approach, CTR increased by 4%.And every dollar spent on use, Dynamic Yield earned $ 6.5 in revenue.
Previously, Lamoda segmented users bylocation and recommended clothing appropriate to the weather conditions. Now recommendations are based not only on geo, but also on purchase history, user behavior, preferred brands and products.
Lamoda divided users by 160microsegments and prepared personalized coupons for each segment. Compared with the previous discount campaign, this has increased conversion, the average income per visitor and revenue per session.
Lamoda launched personal banners on whichclothes and accessories of the color that the user preferred during the search were displayed. When clicking on the banner, the user saw clothes of his favorite color, sorted in the order that he usually prefers when searching.
Thanks to the use of AI, Lamoda increased revenue per session by 8%, andgross profit rose $ 15 million.
Ready solutions, quick effect.
Examples: Aviasales, Weblium, Sephora and Lamodaprove that the use of artificial intelligence helps companies grow significantly, sometimes in a short time: from several months to a year. Moreover, some indicators would never have been improved without the introduction of AI.
You can start experimenting with AIfast. At the initial stage, most likely, the strength of the developers you already have will be sufficient. Search GitHub for developments that can be customized for your company, see if using a completely off-the-shelf third-party product would pay off, and try implementing at least a small idea to see the results. Surely they will impress you so much that you will continue experimenting with AI.