Are Recommendation Systems the Secret to Increased Sales and Revenue?
Amazon Shopping. We’ve all been there, just about to checkout…when all of a sudden a list of related products labeled as “People who bought this, ALSO bought this” appears. The next thing we know, you’ve just added two more products to your shopping cart.
THAT is a Recommendation System.
Recommendation Systems are responsible for dramatically increasing sales and revenue by using two very effective selling techniques; Up-selling and Cross-selling .
According to a McKinsey Report, 35% of all Amazon’s transactions come from their Recommendation System.
The good news is;
- Small Businesses: Small Businesses, Ecommerce and Enterprises can also benefit by incorporating this powerful AI (Artificial Intelligence) tool into their businesses.
- Services: Recommendation Systems are not limited to the selling of just physical products, but are also excellent for selling “Services” such as Health and Wellness Services, Repair and Maintenance Services, Medical and Dental Services, Legal and Financial Services, Energy and Power Services, Beauty and Anti-Aging Services, and Restaurant and Hospitality Services, etc
Here at QAS (Quantum AI Strategy) our specialty is Recommendation Systems. Call today for your FREE Strategy Consultation (480-418-0033).
The remainder of this article is technical deep-dive and will focus on the key motivations that make Deep Learning (Artificial Intelligence) a great tool for the recommendation of additional products and services for purchase.
The main topics of this article are:
- Recommendation Systems Introduction
- Recommendation Systems as a Machine Learning Issue
- Content-Based Filtering
- Collaborative Filtering
- Hybrid Systems
- Cold-start
- Why use Deep Learning for Recommendation Systems?
Recommendation Systems Introduction
A Recommendation System’s main objective is to filter the content that is delivered to the user, try to predict the “grade” or “preference” that the user would give to a content and then make a recommendation based on that preference. Once these recommendations are presented to the user, the likelihood of the user purchasing the recommended content is increased dramatically.
Recommendation systems are a machine learning subarea that aims to make new suggestions that may be of interest to you, based on your background and profile or on choices made by others with similar tastes. It is widely used as a marketing strategy by e-commerce companies, as recommending something in line with the user’s interest increases their chance of acquiring such a product.
The suggestions provided in the Recommendation Systems are intended to assist users in various decision-making processes, as well as which items to buy, which music to listen to, and which news to read.
Recommendation systems have proven valuable in helping users deal with information overload by filtering information and recommending what would be of interest to them.
This way, the consumer experience within the platform becomes more optimized and interesting.
Recommendation Systems can be used to recommend content in different domains, such as books, music, movies, retail, news, etc.
Although there are specifics in each domain, two characters are key, the system users and the content offered by the platform. Thus the basis of Recommendation Systems is the content offered should be filtered in order to generate a personalized and relevant list for each user.
The examples are diverse: Amazon recommends books (or anything in the marketplace), Netflix recommends Movies, Spotify the songs.
In addition, for e-commerce, it has become one of the most powerful and popular tools by recommending products or services according to the habits of users.
You can make recommendations by comparing a user’s preferences with a group of other users. You can also make recommendations by looking for items with characteristics similar to those you have already shown interest in the past. User preferences can be harvested implicitly or explicitly. Implicitly, information is obtained through past purchase options, history of websites visited, clicked links, browser cookies or even geographic location. There is also an explicit way to check preferences using effective feedback, such as grades given for a particular item.
The benefits of a Recommendation Systems for the platform can be varied, such as: improving the user experience within the platform; keep the user logged in longer; increase sales of the products offered or simply all of them at the same time.
Of course, an effective recommendation system is quite complex, its development involves several factors, such as: User profiling and types of information gathering; Use of complementary data for comparison; Distinct filtering techniques; (Re) weight calibration and evaluations; Treatment of the fearsome Gray Sheep, which are characterized by newly discovered behaviors that do not follow the initial pattern; Performance and accuracy, among others.
What Applications Can Recommendation Systems Be Used On?
It is important to evaluate for which applications these systems are viable.
First, they must be based on the items being displayed or offered to users. In another case, the algorithm loses its meaning.
Another important point is that this mechanism is applicable only when there is a large amount of data involved. This is necessary to ensure that the methodology is efficient, since mathematical abstractions are made and the more data, the more accurate the abstraction function, and therefore the more correct the result.
Recommendation Systems and Machine Learning
The system must learn the pattern of user content consumption from past experiences and predict what the user would like to consume in the future.
In general we can divide the different methods / algorithms into different categories of approaches:
Content-Based Filtering
Generates recommendations based on the similarity of content already consumed by the user. That is, it uses the content that the user has already consumed on the platform (read, bought, watched, listened, clicked ..) to generate a profile, so the system searches for similar content that has not yet been viewed by the user and then recommends.
The focus is on item properties. The similarity of a recommended item will be measured by the similarity to the properties of the item that the user has previously purchased or searched for.
The main advantage of this category is that it does not require much user feedback to start recommending something “useful”.
If the first NetFlix-watched movie was “Star Wars”, the system can already recommend the entire universe of movies, series, similar or related, without requiring a very long history of interactions.
There are two main disadvantages to using content-based recommendation:
- The system places the user in a “preference bubble” where everything recommended is similar to what has already been consumed. This long-term preference bubble can lead to user disinterest in recommendations as it generates little diversity in the content presented.
Another point is that depending on the context can be detrimental to the business too, as it does not present new products to the user and the profile may change more slowly than the user’s own preference.
- This problem has more to do with solution modeling. Content-based filtering is performed when filtering similar content, now try to define what content similarity is.
Each context has its particularities and diversities regarding the features that can be used, such as categorical data, free text, numeric, images, audio. The point is that if the similarity function between Content A and Content B does not have well defined, will result in very homogeneous, repetitive or even nothing similar recommendations.
Collaborative Filtering
In the Recommendation Systems based on collaborative filtering the focus is on the relationship between users and items.
The similarity of items is determined by the similarity of their rating by users who have rated the same items, that is, if users have rated items with similar grades, they probably have similar tastes and accept recommendations based on this criterion.
This type of recommendation has positive results in practice, and avoids the problem of repetitive recommendations. One disadvantage is that it requires a large amount of information about the user and their surroundings to work precisely.
In addition, the Collaborative Filtering approach ignores content characteristics and focuses on User X Content interaction. It assumes that the system does not need to know the characteristics of the content, but what content the user consumed to identify which other users had the same consuming behavior.
This way you can “swap” recommendations among similar users when processing the collective.
These are recommendations very similar to “Users who watched Star Wars also watched Lord of the Rings.” This approach solves the two previous problems, takes the user out of the preference bubble, and does not need to define content similarity as features are ignored.
The special attention of this method is in relation to the modeling of the “interaction force” of the User x Content to represent the “How much the user liked the content”.
In practice, this value can be the user’s own rating of the content, the number of times he has accessed the content, how long he has been interacting with, or any other metrics that pass this idea of preference.
The most common way to solve the collaborative filtering problem with ML is to try to infer missing values in an interaction matrix, where input (i, j) describes the “interaction force” or score the user gave to the content.
If the system can ‘predict’ these values it is possible to generate recommendation by selecting the ones with the highest score and the user has not yet interacted. Array Factor Algorithms (ALS, SVD ..) are the most commonly used to ‘predict’ these values in classical approaches, but they are not the only ones.
Collaborative filtering methods also have disadvantages. One of the main ones is the scale of the solution itself, since processing a (collective) matrix of all users with all content is a computational challenge.
An even more damaging detail is that this array is extremely sparse, has many more values missing than filled in, usually less than 3% of the array is actually filled.
Another disadvantage of this approach is the need for a considerable amount of user input and feedback to start generating recommendation. That is, until a new user begins to have recommendations on what to consume, they have to interact with many more items than content-based filtering, which is a problem for platforms with little or no user history.
Explicit and Implicit Preferences
The system user can interact with content in a variety of ways, such as rating, favoriting, sharing, viewing and not interacting, playing…, etc.
The form of this interaction depends greatly on the recommendation domain, for example in the news domain, it could be how long he / she was reading, if the user viewed, shared the text on social networks, etc. The user would hardly rate the news or bookmark it, but in the movie domain such interactions would be more common to perform.
These preferences can be categorized as explicit and implicit. The explicit ones are when the user spontaneously indicates what is important to him and / or the degree of importance, an example would be the user to bookmark the content or to rate it. Implicit data is collected from user behavior within the system, this information may indicate preferences when it is not possible to explicitly retrieve them, an example could be stories viewed on a news site, reading time, percentage of watched video, search terms, etc.
Deciding which information to collect as a user’s preference will depend on the system itself and the domain of the recommendation. Importantly, regardless of the information that will be used as a preference, the goal is always to “model user interest in content”.
This categorization of preferences between explicit and implicit is important in the context of Recommendation Systems, since in many cases the model / architecture is preferably used in one of the categories.
Hybrid Systems
As its name implies, hybrid systems are those that have characteristics or combine the Collaborative and Content Based Filtering approaches. Overall, the results are better than both separately, which is evident when comparing the strengths and weaknesses of each category.
There are several approaches to leaving the hybrid system. The simplest is to unify recommendations generated by separate systems into a single list.
Other more complex approaches add content capability and information to collaborative filtering, or vice versa. They are indeed hybrid algorithms and bring better results.
Cold-start
This category seeks to optimize the recommendation for new users on the platform.
What to recommend when I don’t have user information?
All Recommendation Systems will have new users with little or no personal and consumption history information. A user named Cold-Start is precisely that user within the platform, and depending on the domain the number of cold-start users is larger and more important.
For example, news recommendation has a high rate of cold-start users, usually no one logs in to the platform and each session ends up being a different user, whereas on Amazon or Netflix the user needs to be logged in to buy and consume content.
The simplest way to recommend cold-start is to create generic lists, such as:
- Most accessed content in the last 24hs
- Geolocalized recommendation when it is possible to know the user’s location
- Diverse recommendation lists (best in each category)
Why use Deep Learning for Recommendation?
Just as computational vision and natural language processing were radically impacted by DL, we are currently experiencing the same in the area of Recommendation Systems.
It is not just the amount of academic papers that are growing and advancing state-of-the-art in the last 4 years. Large companies are moving from classic recommendation systems to DL approaches.
Key Points That Make Deep Learning a Great Alternative to Recommendation Systems
There are some features of Deep Learning architectures that make it deep in the Recommendation Systems wave and bringing breakthroughs in the state of the art. The main points are:
- Lots of data. Collaborative filtering techniques are limited by the size of the interaction matrix which are also extremely sparse. DL works very well with large volumes of data, can reduce dimensionality without losing the representativeness of the original information. Convolutional Networks and Autoencoders are widely used in this regard.
- Heterogeneous Data and Features. Depending on the domain, the data representing the content may vary widely, being categorical data, numeric, text, image, audio, etc. Being able to represent all this information in a unified way has a great impact on the similarity of content. DL can easily process and extract patterns in different domains by creating Embeddings, where it is possible to create a vector representation of whatever type of data, this facilitates semantic representation and similarity calculation.
- Dynamic behavior. Depending on the domain, the preferred behavior is very dynamic or short term. Classic algorithms have difficulty extracting patterns when this behavior changes too fast or is not present in the history. DL architectures such as Recurring Networks work very well with this dynamic and especially with the sequential behavior in content consumption.
- Better representation of User vs. Content relationship. The interaction force between these two actors is the basis of collaborative filtering, classical methods end up modeling this interaction in a linear way (ex. Matrix Factorization ..), which limits generalization. Neural networks are known for nonlinear representation of information and can best represent this interaction.
Conclusion
AI (Artificial Intelligence) is quickly changing the business landscape and providing tremendous benefits to those businesses that adopt this powerful technology early. Recommendation Systems have proven time-and-time again to provide a high-ROI and have been rated as one of the best investments a business can make. No more pushy sales tactics, No more embarrassing Up-sell attempts. Recommendation Systems allow you to “Set it and Forget it”. Let the system do the selling for you and see the tremendous increase in your businesses revenue and sales.
Call today for your FREE Strategy Consultation.
QAS (Quantum AI Strategy)
(480) 418-0033.