Наve you ever wondered how does Amazon exactly know what kind of product you need and show you the selection of the most attracting options? Or how can Netflix figure out what kind of movies or series you might be interested in watching? Finally, how does Facebook suggest you “people you might know” and you know the majority of them? All of this goodness is achieved through the intricate recommender systems that gather user information, study it, and turn into an immersive and engaging user experience.
Every user produces a lot of information that, if utilized correctly, can be used to amplify their experience, deepening the relationship between the user and the service by suggesting what the user might need or want at this given moment.
It is essential to understand how to implement recommender systems into the business operation and why they are ultimately the tool that keeps the company on the same page with its customers.
However, first, let’s sort out the basics.
A recommender system is an information filtering algorithm designed to suggest content or products which might be attractive to a particular user.
In other words, it is a more delicate way of bringing user and relevant content together.
Recommender systems became a useful feature due to the necessity to navigate in the sea of content. There is a lot of stuff available online, and many users have a hard time not only with finding something they want but even with figuring out what is it that they want in the first place.
In essence, anything can be recommended or suggested - shoes, clothes, places, films, applications, browser plugins, memes, music, blog posts, communities, even people or their specific skills and so on.
All recommender system does is narrowing the selection of specific content to the one that is the most relevant to the particular user.
Recommender systems are based on combinations of information filtering and matching algorithms that bring together two sides:
- the user
- the content
Two types of user information are in the very center of recommendation operation:
- Explicit data is generated by a user taking a direct action that indicates his/her preferences (for example, ranking the particular product)
- Implicit behavioral data (i.e., user behavior itself) forms the majority of the user profile (for example, what kind of products user views the most)
Overall, the recommender engine operates like this:
- Gathering user data;
- Finding patterns in user behavior and actions;
- Extracting valuable insights;
- Calculating probabilities;
- Comparing them with the available item inventory;
- Presenting the most plausible matches.
Data Mining algorithms handle the first three phases of the operation. Then this raw material is shaped into the user profile which forms the backbone for the recommendations.
Subsequent results are incorporated into the user experience in the form of handy suggestions of various kinds. In addition to that, the machine learning element upgrades the system on the go.
There are three types of recommendations:
- Personalized content feed - the most basic approach. It involves shaping the delivery of the content on the platform according to specific settings. For example, the "home" and “subscriptions” pages on YouTube. In this case, the feed is curated by the platform and the user in equal measure. In one hand user selects what kind of content he would like to see, on the other, the system is looking for similar material to suggest.
- On-site recommendations AKA everything that happens on the platforms. It can be “you might also like” suggestions or “people are also buying” column or some other types of handy notifications (it all depends on the context in which content is suggested);
- Off-site recommendations - suggestions delivered via different channels. It can be a notification on mobile applications or email newsletter or messages from the social media chatbots. These recommendations are often grouped similarly to the on-site ones and coupled with special offers.
Now let’s look at the types of recommender systems.
Let's look at five basic types of recommender systems.
It should be noted that the recommender system is not limited to the use of one particular type and can combine any number of different types depending on the requirements of the specific business operation.
User's feedback is the backbone for the basis for further suggestions in this type of recommendation engines.
The collaborative recommendation is based on the history of user interactions with the platform.
Here's how it works:
- First, the system aggregates the user output - various kinds of search history, ratings, comments, and recommendations of products or pieces of content in a big dataset.
- Then, it compares the output of different users for specific products, finds common elements, and calculates matches between different pieces of content.
This type of recommender system is common in the eCommerce marketplaces. For example, a variation of collaborative recommendation system algorithm is currently used on Amazon.
Overall, the collaborative system is a relatively simple way of making relevant suggestions to the customers. On the other hand, it is a good way of understanding which products are preferred by the users and to what degree in comparison with the other products.
The content-based recommendation goes in the opposite direction from collaborative systems. Instead of focusing on the users' behavior, the content-based recommendation is built around the item inventory (products, content) and attribution comparison.
In this case, if the user is looking for IBM “Think” computers, the system will likely suggest laptops of similar size and tech specifications.
Keywords that describe items lay the foundation for the suggestions and each product usually has more than one keyword to make the matching easier and more precise. These keywords coupled with the user activity form the scope of the product recommendation.
This type of recommender engine is widely used in niche eCommerce stores (Discogs and Artsy use this approach) and also on content aggregation websites with extensive selections of specific content to work around (such as Mashable and The Next Web).
The demographic-based system provides suggestions based on the characteristics of the specific audience segments.
This type of recommendation engine takes available user data (age, gender, location, etc.), classifies it into specific audience segments and then puts in a bigger picture to fill the gaps in the data.
Demographic-based suggestions are widely used on content-aggregation websites and in general eCommerce marketplace. Usually, this type of recommendations provides a background operation in case if there is no other information available.
Demographic-based system is one of the simpler types of recommendation systems that require a limited set of data to deliver broad suggestions. As such, it is less dependent on user data. However, to make it work, this system requires full-on market research as a foundation.
The utility-based recommendation is the one that tries to calculate the usefulness of the particular product according to the expressed preferences of the users.
It is trickier to adjust than the others due to numerous additional elements in the equation.
To calculate the usefulness of the product you need:
- the correlation of the search query
- comparison with the similar considered products
- Product availability
- vendor’s ranking and other relevant elements
Utility-based recommendations require a broad scope of user information available to provide engaging suggestions.
As such it is used for suggesting niche products on multi-purpose eCommerce marketplaces such as Amazon and also on the niche stores for hardware and other products.
The knowledge-based system took a deep dive into the user behavior to calculate the suggestions out of recorded interactions and assumed needs and preferences.
In contrast to the other types of recommender systems, knowledge-based calculates matches and possibilities and attempts to predict the most potent ones.
To do that, the system:
- takes available user information;
- processes it through a combination of predictive and prescriptive machine learning algorithms.
- assesses how a specific product meets user preferences.
The result is a more accurate suggestion with a more substantial potential of getting a conversion.
It is always a good thing when the user finds what he is looking for and also a little bit more. The latter part matters because it leaves a chance for a continuation of the experience. That’s what recommender systems are so good for business.
Content discovery is the reason why recommendation system algorithms are a thing. The suggestions and personalization eliminate the boring parts of content discovery (such as sifting through the stuff the user has no interest in) and concentrate the user's efforts on consuming the content and getting more of it.
It can be presented as an on-site suggestion or a distilled selection through the other channels (email newsletter or social media chatbot).
On the whole, the recommender system perpetuates a loop incited by the user. What started as a single organic discovery of a piece evolved into a narrative of its own by suggesting new pieces that move the user further down the rabbit hole on the particular website.
An excellent example of this is Spotify. Their suggestions are aimed not only at satisfying the user’s direct queries but also recommending something user will be interested in further exploring, which makes the person use the platform more.
Users are not the only ones who benefit from the recommender system's implementation.
On the other side of the system, your company gets valuable audience insights and make the whole operation even better fit for the target audience’s needs and demands.
What is even more important is that these insights are dynamic. As such you will be able to assess the performance and adjust it on the go without the fear of missing the target.
Here’s how it works:
- The system provides users with suggestions.
- Users react to suggestions (positively or negatively - basically follow or don’t follow the lead.)
- The results of the interactions are assessed, and the system is automatically adjusted.
In addition to the gradual upgrade of the system, this information also provides a significant competitive edge in audience research. The marketing department will appreciate this information.
The direct aftermath of delivering relevant content and gathering audience insights is user engagement and retention AKA the most important thing.
Here’s how it works: a constant stream of adaptable suggestions allows to maintain the user interest consistently.
The train of thought behind this kind of engagement is simple: if the particular resource delivers the audience what they want and a little bit more and it all makes a good experience - why not use it more to get more of that positive experience?
As a result, users keep returning to your website for more and regularly checking out for updates through different channels.
Product Recommendation is one of the principles of commerce in general. Suggesting products that might be interesting for the particular customer can trigger the purchase sequence. In e-commerce, this process is handled by the recommender engines.
Structure-wise, e-commerce recommender systems are oriented at:
- assessing the product ratings
- checking the purchase stats
- comparing them with the user behavior
- making relevant matches for the suggestions
To do that e-commerce recommender engines are using a combination of collaborative and demographic techniques often mixed with content-based methods.
The combination of different types of recommender systems allows to fill the missing data more efficiently and perform a more broad comparison of user preference and available product inventory.
Such an approach is one of the reasons why Amazon retains such a dominant position in the eCommerce industry.
Content aggregation platforms seem to be a natural fit for the recommender systems. After all, when there is too much content to consume, the user might feel lost.
In this case, suggestions are one of the more effective ways of navigating in the sea of available content and keeping the users on-site for a longer time with the relevant content.
One of the advantages of content aggregation platforms is the flexibility of content. It can be filtered in a variety of ways using multiple configurations of recommender engines.
For example, YouTube is using a combination of the following types of recommendation engines:
The insights extracted with several algorithms allow more precise content targeting which guarantees retention and further activity on-site. That, in turn, contributes to the improvement of the recommender system algorithm via machine learning.
On the other hand, there is Netflix Recommendation Algorithm, which uses knowledge-based and utility-based approaches with the assistance of a collaborative approach to calculate the most appropriate suggestions for the viewer.
The difference between YouTube and Netflix methodologies is in the following:
- YouTube presents the content moderated according to the user’s expressed interests.
- Netflix suggests the content that will likely fit the user’s expressed tastes.
Spotify uses a similar approach to Netflix as it perpetuates the use of the platform by making intriguing suggestions based on expressed user interests.
It should be noted, that Content Aggregation platforms are often plagued with poorly labeled data which makes large portions of content virtually invisible for the search and recommendation engines.
The relevance and longevity of the social media platforms depend on the nature of connections they help to make. In this case, the use of the recommendation engine is the key to providing consistent user engagement with strong retention rate and everlasting discovery process.
Facebook uses different types of recommendation engines for different kinds of suggestions. To form “people you might know section” - content-based and demographic recommender systems are used;
- To form newsfeed sections, it uses content-based, collaborative and utility-based algorithms.
- To promote companies or places, Facebook adds a collaborative recommendation to the mix as it involves a direct rating of the enterprises.
- The system also applies elements of knowledge-based recommendation to diversify the selection of the suggestions.
LinkedIn is a professional-use oriented platform. One of its prevalent use cases involves HR recruiters searching for suitable candidates for the vacant positions. The role of a recommender system algorithm here is to provide suggestions based on relevant attributes. Essentially, the methodology is similar to the content selection but with a broader range of settings.
The system uses a combination of the following approaches to make better matches:
Recommender systems are one of the cornerstones of the active customer service. They make the whole “giving the user what they need” much easier. In addition to that, these systems are a good way of practical exploration of the needs and demands of your target audience. As such, recommender systems are an indispensable part of the modern customer service.