About Recommender Systems

Recommender systems have become an essential tool for finding the needle in the haystack of the World Wide Web - the information or item one is searching for. Finding the desired item is a daunting task considering the amount of information that is present in the WWW and its databases, and almost every e-commerce application will provide the user with the help of a recommender system to suggest products or information that the user might want or need. Recommender systems are employed to recommend products in online stores, news articles in news subscription sites or financial services like retirement plans. There are different recommendation methods, each one with its advantages and disadvantages. In order to provide individualized recommendations to each user of such a system, personalization is of high importance. In order to be able to personalize the use of a system, users are often modeled by the recommender system. Generally, the user model represents the user's taste, and it is generated, learned, or extracted from the information available about the user. Again, there are several methods to create a user model, which are often characteristic for every recommendation technique. User models can be built on explicit and implicit data. While explicit data refers to the data directly provided by the user in the form of, e.g. a questionnaire, implicit data refers to data provided implicitly by e.g. the user behavior. Often user models are created using both types of information.

There are three main methods commonly employed in recommender applications: collaborative filtering, content-based filtering, and knowledge-based methods. Collaborative filtering, the most efficient and popular approach, is based on the idea that people who share the same taste will like the things some similar-minded person preferred. For capturing a user's taste, ratings of the available items are employed to represent the user. Ratings are typically numerical values standing for the degree of utility or subjective value of an item. By computing similarities between rating patterns of users, the rating of a user for an unseen item can be predicted, and eventually recommended. In that process, a neighborhood of similar users is detected, and their ratings will determine a prediction value for an item the user has not yet rated. Since ratings are used to calculate recommendations, a collaborative filtering system needs many ratings in order to work properly cold-start problem. Once enough rating data is available, the technique is very efficient and increases in accuracy as more information is available. Further, the probability increases that a user with exquisite and rare taste receives satisfactory recommendations (niche-finding). Also, a new user that did not provide any ratings yet, cannot receive personalized recommendations new user problem. Another disadvantage of the approach is the inability to recommend a new item, since nobody rated that item yet new item problem.

In content-based techniques users are not compared to users, but to items. A user model is learned on basis of the content, i.e. the features or the properties of the items a user liked. New items can be recommended that are similar to the ones previously liked by the user. The user is modeled using a feature vector, representing his interest in form of key features of the items he likes. Since this technique is dependent on information about the items in order to provide recommendations, data has to be extracted and converted into a machine processable way. Because this is done straightforwardly with digital documents or texts, content-based techniques are mainly applied in the domain of document recommendation systems, like webpage, news article or book recommender systems. Applying it to other domains can be very costly since all the information about the items has to be represented somehow. The cold-start problem is also present in content-based recommender systems, as well as the new user problem. The new item problem is avoided since content-based filtering does not depend on how others rated that item in order to recommend it.

Knowledge-based recommendation systems apply techniques from Artificial Intelligence and especially from knowledge representation and reasoning in order to provide recommendations. By using knowledge about users and items a knowledge-based recommender system will reason about what items meet the user's requirements. That knowledge can be encoded in the form of e.g. rules and saved in the knowledge base. Different from the other approaches, the user's individual taste is not in focus but his specific requirements. These needs are inquired in form of a dialog with the system or by the use of a questionnaire. On the basis of the knowledge about the items and these requirements, a reasoning process will find the best matching item. The reasoning process in knowledge-based recommender systems and also the knowledge base can be implemented using different techniques. Logic programming languages are often employed for their elementariness in representing knowledge and defining reasoning mechanisms. An advantage of knowledge-based recommender systems is the easiness of deriving explanations from the reasoning process. Every recommendation can be traced back to the requirements of the user and the knowledge used to infer the recommendation. The need for knowledge engineering and maintenance of the system, when e.g. new items have to be entered, are major drawbacks of these systems and often do not allow employing them in very large domains.

Hybrid recommender systems are using more than one technique in order to overcome technique specific disadvantages. This merging of techniques can be carried out using one of several interaction types that define how the two techniques will work together. One technique can, for example, be used to produce recommendations, and another technique will refine these recommendations. This interaction is called cascading. Another hybrid method is switching. In a switching hybrid recommender system, some criteria is used to switch between techniques depending on the situation. Hybrids using e.g. the collaborative filtering and the content-based method can avoid the new item problem, while retaining the efficiency and niche-finding capabilities of collaborative filtering.

Recommender systems that are employed in online applications with many users and large amounts of items, have their priority set on performance. The aim is to provide accurate recommendations to each user, while maintaining efficiency. The larger the system the more important if the factor of efficiency. This often prevents to employ an expressive user model. In some systems, the user does not even get modeled individually but by means of a group definition. This can highly impact the accuracy of recommendations. Further, recommender systems barely provide a way for the user to specify individual preferences explicitly. Most systems only offer feedback mechanisms where e.g. already presented recommendations can get refined or where the user can exclude items from the recommender process. Moreover, the explanation mechanisms are quite poor. Due to the statistical and similarity methods used in e.g. collaborative filtering, it is very difficult to provide an explanation once a recommendation is calculated. While knowing that the ratings of items are the initial factors of a recommendation process, the process itself remains a "black box". Especially the involvement of other user behavior makes it a complicated task to explain recommendations in a user understandable way. And the information provided in present systems can make the user undervalue the power of the such systems by believing that it only infers recommendations based on primitive similarity calculations. Further, this situation can make the user lose trust. systems that are employed in online applications with many users and large amounts of items, have their priority set on performance. The aim is to provide accurate recommendations to each user, while maintaining efficiency. The larger the system the more important if the factor of efficiency. This often prevents to employ an expressive user model. In some systems, the user does not even get modeled individually but by means of a group definition. This can highly impact the accuracy of recommendations. Further, recommender systems barely provide a way for the user to specify individual preferences explicitly. Most systems only offer feedback mechanisms where e.g. already presented recommendations can get refined or where the user can exclude items from the recommender process. Moreover, the explanation mechanisms are quite poor. Due to the statistical and similarity methods used in e.g. collaborative filtering, it is very difficult to provide an explanation once a recommendation is calculated. While knowing that the ratings of items are the initial factors of a recommendation process, the process itself remains a "black box". Especially the involvement of other user behavior makes it a complicated task to explain recommendations in a user understandable way. And the information provided in present systems can make the user undervalue the power of the such systems by believing that it only infers recommendations based on primitive similarity calculations. Further, this situation can make the user lose trust. systems that are employed in online applications with many users and large amounts of items, have their priority set on performance. The aim is to provide accurate recommendations to each user, while maintaining efficiency. The larger the system the more important if the factor of efficiency. This often prevents to employ an expressive user model. In some systems, the user does not even get modeled individually but by means of a group definition. This can highly impact the accuracy of recommendations. Further, recommender systems barely provide a way for the user to specify individual preferences explicitly. Most systems only offer feedback mechanisms where e.g. already presented recommendations can get refined or where the user can exclude items from the recommender process. Moreover, the explanation mechanisms are quite poor. Due to the statistical and similarity methods used in e.g. collaborative filtering, it is very difficult to provide an explanation once a recommendation is calculated. While knowing that the ratings of items are the initial factors of a recommendation process, the process itself remains a "black box". Especially the involvement of other user behavior makes it a complicated task to explain recommendations in a user understandable way. And the information provided in present systems can make the user undervalue the power of the such systems by believing that it only infers recommendations based on primitive similarity calculations. Further, this situation can make the user lose trust.