If you understand the mechanism of the recommendation engine and AI (machine learning) algorithm, you can maximize the effect.
In this article, we introduce seven types of algorithms for recommendation engines.
In the second half of the article, we explain how to create an algorithm that suits your company, so please read it to the end.
Mechanism of seven major types of AI (machine learning) algorithms for recommendation engines
Recommendation engines have seven main types of algorithms.
- Rule-based recommendation
- content-based filtering
- personalized recommendations
- collaborative filtering
- Image/audio analysis recommendation
- Graph AI technology
- hybrid filtering
By understanding the recommendation mechanism, you can choose the tool that best suits your company and maximize its effectiveness.
Let’s take a closer look at the seven algorithms.
The most commonly used recommendation method is “rule-based”. In the rule base, operators set fixed rules and make recommendations according to them.
- Recommend Hina Arare during the Hinamatsuri season
- Recommend coffee to someone looking to buy cookies
As mentioned above, we anticipate the customer’s behavior and preferences and enact rules that we think are optimal.
It is an effective method when you want to recommend a specific product, but it is not based on the data of customers who are actually browsing the site, so it is not clear whether it is the best proposal for the customer. need to be careful.
Content-based filtering is a recommendation method that recommends items with a high degree of similarity according to the genre and attributes of content and products.
- Recommend a different white skirt to someone who sees a white skirt
The high accuracy that can recommend items that are close to what the customer is interested in is attractive.
Just pay attention to the following three points.
- It is difficult for customers to come across unexpected items
- Customer behavior and purchase history are irrelevant
- Attributes must be sorted in advance
“Personalized recommendations” recommend items that are considered to be close to the customer’s tastes and hobbies, although they are not based on behavior or purchase history.
- Recommend a new bag from the same brand to someone looking at a brand A item bag
- Recommend whitening care items to people looking at whitening lotion
Personalized recommendations are not only set based on the persona’s tastes and hobbies, but there are also cases where recommendations are made based on a simple survey of the customer themselves.
Up to this point, we have introduced methods for following certain rules and recommending items that match the attributes of the content and persona without referring to the customer’s behavior history or purchase history.
Collaborative filtering, on the other hand, is an algorithm that utilizes customer behavior and purchase histories. There are two types of collaborative filtering: item-based and user-based.
Item-based collaborative filtering suggests items that are likely to be purchased together with the product.
- Propose the purchase of protective glass when purchasing a smartphone case
User-based collaborative filtering suggests recommended items based on the behavior and purchase history of users who have purchased the same product.
- “People who purchased brand A eyeshadow also purchased brand B lipstick,” said the purchaser himself, who did not consider the brand or item, but purchased it among those who purchased brand A eyeshadow. Recommend high-value items
Collaborative filtering is based on actual behavioral history and purchase history, so it has the advantage of high CVR.
Image/audio analysis recommendation
Until now, recommendations were made based on the attributes of products, items, and content registered on the site, but in recent years, it has become possible to make recommendations by analyzing images and sounds.
- Recommend alternative songs with similar signals to the user’s favorite songs
Graph AI technology
“Graph AI technology” is attracting attention as a mechanism that applies deep learning to graph data and finds hidden relationships.
A user or an item purchased by a user is called a “node”, and the relationship with a node is called an “edge”. It is a method to understand not only the characteristics of the node itself, but also the actual situation more three-dimensionally and grasp new characteristics.
This method is also used in Uber Eats and Pinterest, and it can be said that it is an effective algorithm for large-scale sites and sites where the number of products is increasing.
A method called “hybrid filtering” combines multiple algorithms introduced so far.
In hybrid filtering, content-based recommendations such as “For those who are looking for a white skirt, I would like to introduce another white skirt as a related product” and collaborative filtering such as “Introducing popular brand items in their 20s” are performed at the same time. is also possible.
Hybrid filtering should be considered if recommendations can be made from various perspectives, such as “genre of items and content” and “user attributes”.
Types of recommendation engines and how to make them
There are mainly two types of recommendation engines: ASP type and open source type. We will discuss the features, advantages, and disadvantages of each.
ASP stands for “Application Service Provider” that provides web application services on the network.
The greatest advantage of the ASP type is that it is easy to start using immediately after installation. It is a nice point that it can be introduced in a much shorter period of time than developing from scratch, and the cost performance is good.
However, it is necessary to understand that it is difficult to add or customize your own functions. If there is a function you want, look for a tool that is provided or develop it.
open source type
The open source type is a method of developing a recommendation engine based on publicly available source code.
The biggest advantage of the open source type is that it can be developed freely. If you want to add your own functions, it is a good idea to consider the open source type.
However, it is necessary to be aware that engineers who can develop are essential, and that many of the open source types are made overseas.
How to choose a recommendation engine algorithm that suits your company
There are two points to consider when choosing a recommendation engine that suits your company.
- Examination of algorithms suitable for your company
- Analysis of AB tests and reports
Examination of algorithms suitable for your company
When choosing an algorithm for a recommendation engine, it is important to understand how it works and consider what kind of algorithm is suitable for your company.
- Rule-based recommendation because customer behavior data and customer data are scarce
- We want to use content-based filtering There are a large number of products and customers use them frequently, so we want to propose products that suit customers through personalized recommendations.
Consider the algorithm that suits your company, as described above.
Instead of choosing just one algorithm, you can combine multiple algorithms as needed.
The cost increases by using advanced technology and combining multiple algorithms. Consider the cost of introducing the tool and the effect when the accuracy is improved, and choose the tool that suits your company.
Analysis of AB tests and reports
The introduction of the recommendation engine does not end there. By conducting AB tests and analyzing reports, it becomes possible to make more effective recommendations.
By considering the ease of operation, such as whether the recommendation engine has an AB test function, what kind of content is output in the report, and whether the necessary information can be extracted in an easy-to-read manner, it will be effective in the long term. will be easier to feel.
Understand the algorithm and mechanism of the recommendation engine and choose the system that suits your company
Recommendation engine algorithms range from rule-based algorithms that follow preset rules to collaborative filtering that utilizes data and algorithms that use AI technology.