Advanced Analytics and Deep Learning Models. Группа авторов
systems that collect information on multiple criteria. The basic working of recommender system is to predict accurate recommendation for a particular user. The recommender system or single-criteria recommender system explores only one criteria and give recommended result. This is the first ever recommender system concept. But for real-world problems, we cannot predict recommendation list by exploring only one criterion at a time. It will give false prediction. So, the MCRS concept comes in the field. These kinds of recommender system can explore multiple-criteria at a time and can give excellent accuracy (Figure 3.1).
Figure 3.1 Working principle of MCRS.
Figure 3.2 Phases of MCRS.
Recommender systems are widely used in e-commerce systems and movie industries and each and every sector. Suppose if we used amazon. com and buy a product, then before check out it shows similar kind of product as add on. This list of items is predicted by amazons very own recommender system. Similarly, if we use Netflix, then we can see that it always recommends new movies and web series to us. This prediction is based on generally two categories, on the basis of our previous choice and other one is on the basis of out Netflix account details. That is how recommender system generates a list which is most suitable to the user.
Every recommender system goes through three types of phases. Those phases are modeling phase, prediction phase, and recommendation phase.
Figure 3.2 explains the different phases of a recommender system. Now, we will see the each phases and their significance.
3.3.1 Modeling Phase
In modeling phase, its focus on preparing the data will be used in next two phase. As we can see in the diagram, it is also divided in three cases. First step is to build a ratio matrix. The rows of the matrix contain the name of the users, and columns contain the items and each cell contains rating, which by the user for a particular item. Now, it generates a user profile. This profile explains the preference of a user. It is mostly a vector and every user has their own private profile of preferences. In the third step, it generates a profile for the items which contains the features of the items.
3.3.2 Prediction Phase
It is the second phase of recommender system. The main objective is to estimate the rating or score of unrevealed or unspecified items for every client. This process is done by a utility function based on the extracted data which is provided by the modeling phase. Different filtering techniques are diagrammatically shown in Figure 3.3.
Figure 3.3 Filtering techniques of MCRS.
3.3.3 Recommendation Phase
This the third phase of recommender system and also extension prediction phase. In this step, various methods are used to hold up clients’ choice by predicting the most acceptable items. As per the user’s interest, new items are recommended in this step.
These are the most three important phases of a MCRS. We use different kind of approaches in MCRS to predict good results. The most important and widely used approaches are content-based filtering, CF approach, and knowledge-based approach. We also have a hybrid approach. Now, we will take a deep look about all these approaches of MCRS.
3.3.4 Content-Based Approach
It generates the suitable recommendations for a client that depends on his previous behaviors. It analyzes the user’s previous history like what liked, bought, or watched and accordingly it predicts. It generates a user profile for every user based on their previously selected items and recommends items to him based on similar features items which he liked before. It does not compare his preferences to the users to characterize each user. Content-based filtering approach is divided in three steps which are item representation, learning the user profile, and recommendations generator. In the item representation step, the information or the description of item is extracted to create item’s characteristics. It produces the structured item’s representation. In the next step, a user profile is generated. This user profile is based on the previous behavior such as liking or disliking, the rating or by writing some text comment given by the user for a particular item. This step is known as learning the user profile and the last step is recommendation generator. In this step, a list of recommended items is generated and compared it with the item’s features of the client’s profile. The item that is suitable or most likely is added to the prediction list [1].
This method has been executed in various domains like textual details such as websites, news, and articles and also used for recommending activities such as tourism, travel, TVs, and e-commerce industries. This method works very efficiently if the items size is moderate. This approach relies on content or characteristics of each item so it gives several advantages like it offers a high level personalization in recommendations; it can make its scale up when number of users increases, which means it is scalable. It can make recommendations with a particular interest of a user and it provided a very good security also. These are some advantages of content-based filtering approach [1].
3.3.5 Collaborative Filtering Approach
This technique is immense popular technique among all the multi-criteria recommender system. It interacts with multiple users and generates the recommendation list. If user1 has similarities in their preference with user2, then the item which is recommended to user2 will also be recommended to user 1. The hypothesis behind the following approach is that the clients agreeing other clients in the past will also agree in the future. For a new item, the relationship with user is determined by other users’ review. We can represent as user terms matrix where each cell of the matrix represents the ratings given by clients for a particular item [1].
CF can be divided into two classes: model-based and memory-based. The memory-based approach is a kind of heuristic algorithm. It estimates the item’s rating which depends on another client’s ratings. It can also be classified into two methods: item-based and user-based. The other one, memory-based approach, recommends items based on the similar interests on other users. It analyzes the behavior of other clients like they purchased or liked or viewed before and then recommend the product to this client [1].
CF approaches have many advantages compare to all other filtering approaches like, sometimes, novel and unfamiliar items are recommended, it is very suitable and flexible in various domain, and it does not need to analyze the contents of a particular item [1].
These are some of the advantages of CF in MCRS.
3.3.6 Knowledge-Based Filtering Approach
This is comparatively a new approach than other two approaches. This method is used in those cases where both collaborative- and content-based approaches failed or cannot work properly. The situation happens when there is not enough ratings or reviews are at hand for a particular item for the recommendation process. It is generally happening for those that are hardly ever purchased like houses, cars,