This chapter has reviewed existing studies related to user profiling and recommender systems. The goal of user profiling is to collect information about the subjects or topics in which a given user is interested. User profiles are the main source of information through which personalisation systems can learn about users’ interests or preferences. As the literature review indicates, information about the interests of users can be obtained either by asking users direct questions (commonly called explicit feedback) or by indirect means (referred to as implicit feedback).
Most personalised recommendations have focused on achieving recommendation accuracy. A major challenge within user profiling is how user profiles can be constructed that accurately reflect users’ preferences. This includes the question of how to obtain new users’ preferences if the available information is limited. Besides using ratings data, additional information about users and items can be harnessed to enhance the performance of recommender systems and address problems, such as the cold-start problem and the sparsity problem. Recently, the usage of item taxonomy has been emphasised by many researchers seeking to improve recommendation performance and alleviate the cold-start problem. Taxonomy information is expensive to gather, but it is considered to be better structured and more widely applicable than standard item content information. In addi- tion, a few studies suggest the use of a language model with the information retrieval (IR) method to solve recommendation-related problems, as proposed here.
The following chapters will extend existing knowledge by offering effective ap- proaches to using item taxonomy information to obtain the personal preferences of a user when user rating data and profile information are limited. Furthermore, new methods based on the use to taxonomy information for improving performance in recommender systems will also be examined in this thesis.
ACQUIRING USER INFORMATION NEEDS
BY CONCEPT HIERARCHY
Acquiring information about user preferences is an important task because it is the ini- tial step in building personalised recommendations. Every recommender system has to develop a user model or a user profile that contains the personal preferences of the user. The challenge of making personalised recommendations is the effective acquisition of user preferences (or needs) when there is limited personal data about users, such as when new users have only rated a small number of items [19, 102]. The system cannot generate accurate recommendations for these users, which is commonly called the cold- start problem.
A variety of techniques to solve this problem have been developed, such as discov- ering user preferences from user interaction with specific items [7], maintaining a user profile, building a model of user preferences to identify the needs of individual users [48], or using additional information about users, such as gender, age, and geographical location and items, such as genres, products categories, keywords, and product descrip- tions [32, 57, 75, 85]. In addition to using ratings data to capture users’ item preferences, item taxonomic information is another popular textual source of information and provides an alternative data source for acquiring user preferences [67]. Taxonomy information is used widely on e-commerce sites. Item taxonomy contains a set of categories or topics designed by web managers or website experts based on their personal experiences. Items
can be described or classified according to sets of categories or topics.
This thesis provides a unique method to solve the cold-start problem, an alternative way to deal with uncertainties when acquiring user preferences. Instead of utilising user ratings, this thesis proposes a novel concept hierarchy model to determine user interest in items according to taxonomic concepts that the user uses and which relate to the items in question. The user preferences and item content can be represented as a user concept hierarchy and item concept hierarchy, respectively. This chapter proposes using item taxonomic information as domain-specific knowledge.
3.1
PROBLEM DEFINITION
Most recommender systems make recommendations based on users’ item preferences, which are extracted from user ratings data. The performance of recommender systems is diminished when they have only a few ratings or insufficient personal data. The recommender systems cannot surmise the relationship between users and items when there is insufficient information, such as the first time a new user visits a system or when a new item is added and has no ratings data available. The challenge in this context is how to obtain the personal preferences of new users when the systems have insufficient information with which to generate high-quality personalised recommendations. This requires the development of new techniques to alleviate the problem and produce better recommendations.
In addition, user preferences involve fuzzy information or knowledge. Fuzzy knowl- edge can be defined as information or concepts that are vague, imprecise, uncertain, or ambiguous in nature. Most users do not know how to describe exactly what they want or choose the right information to fits their personal needs. Consequently, the systems cannot
accurately capture user preferences or needs, which leads to poor recommendations. Therefore, to profile users accurately, we must discover how to help them to accurately identify their own needs from the fuzzy information in their minds.
In addition to using ratings, other factors that can be used to learn user preferences, such as their topics of interest, the order of items in the recommended item list, or the users’ taxonomic topic interests [44, 67, 111, 121]. To address the problem, additional information about the users or items must be exploited [75, 85]. This thesis demonstrates how to item taxonomic information to identify user information needs and preferences to support user profiling and to help deal with the aforementioned recommendation-making problems.