For the efficient personalization of offers using AI, correct business goals are needed. This is what Denis Arismiatov thinks, the Head of Marketing at Homstersonline real estate marketplace.
On June 4, Denis and his colleague Volodymyr Rudyi will make a presentation at AI Conference Kyiv explaining how machine learning can help to personalize offers that websites make to visitors. Therefore, we decided to find out who may find such AI solutions useful and how a business can get ready for their implementation.
Interviewer: AI Conference Kyiv (AICK).
Respondent: Denis Arismiatov (D.A.).
AICK: Hello, Denis! At AI Conference Kyiv, you will present a case study of using ML to personalize offers in real estate. How do you think, which industries may find machine learning useful to personalize offers for their customers?
D.A.: I think that this technology may find use in all fields that have many users and a product to offer. Especially in cases when the website offers a huge product catalog and visitors occasionally return to the website to make another transaction. Most often they are:
- online stores;
- websites for watching videos (for example, Megogo);
- cinemas (for communication with customers using messengers or emailing);
- classifieds (for example, OLX).
Most modern websites are using personalization in recommendations in some way. It is just a question of how efficient recommendations are, whether websites estimate the efficiency and runA/Btests. This is required to see the difference in every new model and improve results gradually.
Personalization helps both the buyer and the seller. The buyer finds the wanted product and accessories quicker, and the seller increases the return on marketing investment (ROMI).
The benefit is that the buyer(in case the rest of sales chain functions seamlessly) is happy with the online store and the seller gains a loyal customer and the increase in the number of users that return to the website.
ML-based recommendations can be also used when there is no great number of goods on the website. For example, banks can communicate with customers and offer them products or individual offers. Here one can rely on user behavior, how often customers use credit cards, and other actions.
AICK: Why artificial intelligence can cope with the task of personalization better than a human?
D.A.: Most likely AI will not cope with it better but much quicker. For a person to be able to recommend something to another person, this person has to study loads of data about the behavior of a specific user and other similar users. It is necessary to ask tons of questions and there is no guarantee that answers will be fair. And of course there is always a human factor – you can miss something.
A well-trained ML-based model can generate an appropriate offer in a matter of seconds when it has access to correct data. And when a specific user returns to the website the next time, the model re-estimates results and provides updated offers. It relies on new input data provided by the user during his visit.
AICK: Let’s suppose that I own a large online store. What should I do to get ready for the implementation of the AI solution that would personalize offers for my customers?
D.A.: You should clearly set up the goal you want to achieve. For example, you want to increase the sales conversion rate generated by users that visited the catalog from 1% to 1.5%.
Further, you should turn to a company that has experience in the implementation of such solutions or consult with a trusted expert that has analogous experience. Otherwise, you risk losing tens or even hundreds of thousands of dollars with no effect.
You should employ at least one data science specialist that would build recommendation models based on ML. The company or an experienced person outside your firm will be able to consult and train the new member of your team. Of course, if you hire a strong specialist that has already developed recommendation systems for other projects, most probably you will not need any consultancy services from the third party.
Then you have to set up the collection of required data about user behavior on the website: what pages they visit and where they come from, how they interact with navigation on the site, what they see, what they klick, what they buy. Arranging the correct collection of data is almost the most important step in machine learning.
When you accumulate the required volume of data to build models for recommendations, your data scientist will be able to test different solutions, for example, collaborative filtering, matrix factorization, and others.
After you train the model, you will need the help of your back-end developers. They should implement the recommendation model in the product catalog (most probably the code will be written in Python).
Then to assess the efficiency of ML-based model, you will have to launch A/Btesting. One part of visitors will see a standard version of your catalog, and the other – a personal version based on ML. The criterion that you have initially chosen to estimate efficiency will be the main characteristic. In such a way, you will understand whether the new machine learning-based model works efficiently.
Further goes the stage of optimization and improvements. This constant process never ends. If the first version of the recommendation system shows a better result than the standard version of your catalog, it will become the main one. During the next A/Btesting, it will compete with the updated model prepared by your specialist.
AICK: What difficulties will the company face during the implementation of such a solution?
D.A.: From my experience, the most difficult things are:
- to find qualified and experienced specialists to act as consultants;
- to find qualified and experienced specialists to become a part of the team;
- to set up data collection;
- to provide stability and consistency when collecting data;
- to implement the recommendation system in the production version of the website and run А/B testing;
- to measure efficiency.
Besides, you should always remember that if the main evaluation criterion improves (for example, the sales conversion rate increases), it is quite possible that other metrics will show negative dynamics (for instance, retention rate). You should monitor these things and conduct many-sided analytics.
AICK:How fast will the AI solution for the personalization of offers payback? What does it depend on?
D.A.: Timeframes can be very different here. The solution can pay back in three months and it can happen that it will not pay back at all. Nobody can guarantee that personalization will help your business.
Much depends on the correctness of goal setting before you start doing something. Therefore, it is important to ask yourself questions:
- What do we want to improve?
- Will we be able to collect the required data?
- Will we be able to allocate resources for developers to set up data collection and implement solutions of the data science specialist?
AICK: In your opinion, what hinders the adoption of such AI solutions?
D.A.: People often do not understand what should be done and where to start. Besides, ML-based personalization requires rather big expenditures that do not guarantee quick improvement of characteristics. Of course, not everyone needs this.
AICK: Please, share what you will be talking about at AI Conference Kyiv.
D.A.: Volodymyr and I will tell why and how to implement recommendation systems on websites. What data collection solutions exist? Why is it so important to collect correct and full data about all users? Through a specific example, we will show the work of the recommendation system based on collaborative filtering. We will also tell about case studies that we received thanks to the recommendation system on Homstersplatforms.
Register to AI Conference Kyiv to listen to the report of speakers.
You'll know more than your colleagues and business rivals.