Recommendation System

According to some known author Recommendation System is a

System that produces individualized recommendations as output or has the effect of guiding the user in a personalized way to interesting or useful objects in a large space of possible options

Burke (2002)

 

A personalized information filtering technology used to either predict whether a particular user will like a particular item (prediction problem) or to identify a set of N items that will be of interest to a certain user (top-N recommendation problem).

Deshpande and Karypis (2004)

 

Recommender systems suggest items of interest to users based on their explicit and implicit preferences, the preferences of other users, and user and item attributes

Schein et al. (2005)

 

Popular Recommendation Engines

Some popular recommendation engines which have been proved as highly profitable:

• the Amazon.com,
• the Netflix.com and
• the Google.com recommendation engines

Amazon.com claims that 35 % of products sales result from recommendations.
About 66 % of movies rented in Netflix.com are recommended.
Google News Recommendations generate 38 % more click-throughs .
 

Recommendation Approaches

The approaches of recommender systems: collaborative filtering (CF), content-based Filtering (CB) and hybrid methods:

• Collaborative filtering algorithms recommend those items to the target user, that have been rated highly by other users with similar preferences and tastes. Websites that provide recommendations in the form, “Customers who bought item i also bought item y“, typically fall under collaborative filtering approaches
• Content-based filtering uses the information derived from documents or item features (eg. terms or attributes). It uses a set of attributes, which describes the items and recommends other items similar to those that exist in the user’s profile. This way, the cold start problem for new items and new users are alleviated, provided that users prefer items that are similar in content to those they have already chosen. However, the pitfall is that there is no diversity in the recommendations. That is, the user gets recommendations that are very familiar to her, since the recommended items are similar to those in her item profile
• Hybrid algorithms attempt to combine Collaborative filtering with Content-based filtering. The combination of content with rating data helps capture more effective correlations between users or items, which yields more accurate recommendations.

Recommendation System Types

Other types of recommender systems proposed (by Burke) in the literature :

• Demographic recommendation, which classifies the users according to the attributes of their personal profile, and makes recommendations based on demographic classes
• Utility-based recommendation, which makes suggestions based on a computation of the utility of each item for a user, for whom a utility function has to be stored
• Knowledge-based recommendation, which suggests items based on logical inferences about user preferences. A knowledge representation (e.g. rules) about how an item meets a particular user need is necessary.

Recommendation Problem

• Object of the decision. That is, defining the object upon which the decision has to be made and the rationale of the recommendation decision.
• Family of criteria. That is, the identification and modelling of a set of criteria that affect the recommendation decision, and which are exhaustive and non-redundant.
• Global preference model. That is, the definition of the function that aggregates the marginal preferences upon each criterion into the global preference of the decision maker about each item.
• Decision support process. That is, the study of the various categories and types of recommender systems that may be used to support the recommendation decision maker, in accordance to the results of the previous steps.

Recommendation Capabilities

• Choice, which involves choosing one item from a set of candidates.
• Sorting, which involves classifying items into pre-defined categories.
• Ranking, which involves ranking items from the best one to the worst one.
• Description, which involves describing all the items in terms of performance upon each criterion.

Recommendation technique used in TEL

 

Collaborative filtering (CF) techniques
Name	Short description	Advantages	Disadvantages	Usefulness for TEL
1. User-based CF	Users that rated the same item similarly probably have the same taste. Based on this assumption, this technique recommends unseen items already rated by similar users.	– No content analysis – Domain-independent – Quality improves over time – Bottom-up approach – Serendipity	– New user problem – New item problem – Popular taste – Scalability –Sparsity – Cold-start problem	– Benefits from experience –Allocates learners to groups (based on similar ratings)
2. Item-based CF	Focus on items, assuming that items rated similarly are probably similar. It recommends items with highest correlation (based on ratings to the items).	– No content analysis – Domain-independent – Quality improves over time – Bottom-up approach – Serendipity	– New item problem – Popular taste – Sparsity – Cold-start problem	– Benefits from experience
3. Stereotypes or demographics CF	Users with similar attributes are matched, then recommends items that are preferred by similar users (based on user data instead of ratings).	– No cold-start problem – Domain-independent –Serendipity	– Obtaining information –Insufficient information –Only popular taste –Obtaining metadata information –Maintenance ontology	–Allocates learners to groups –Benefits from experience –Recommendation from the beginning of the RS
4. Case-based reasoning	Assumes that if a user likes a certain item, (s)he will probably also like similar items. Recommends new but similar items.	– No content analysis – Domain-independent – Quality improves over time	– New user problem – Overspecialisation – Sparsity – Cold-start problem	– Keeps learner informed about learning goal – Useful for hybrid RS
5. Attribute-based techniques	Recommends items based on the matching of their attributes to the user profile. Attributes could be weighted for their importance to the user.	–No cold-start problem –No new user I new item problem –Sensitive to changes of preferences –Can include non-item related features –Can map from user needs to items	–Does not learn –Only works with categories –Ontology modeling and maintenance is required – Overspecialisation	– Useful for hybrid RS – Recommendation from the beginning
6. Decision Trees (C4.5,ID3)	A decision tree represents a set of classifications created from a set of rules. They start form a single classification and branch out based on classification rules mined from the data.	–Easy to understand –High representation power	–Overspecialisation in small datasets –Can become very broad	–Visualize differences of learners from the data –Alternative approach to expert driven ontologies
7.K-Nearest Neighbor (Isodata, Forgy)	Does not build an explicit model instead exams the categories of the K-most similar data points. K-means is often used in TEL recommenders to compute similarity of vector-based approaches.	–Simple approach only two parameters to select –Robust to noise –High representation power	–Difficult to select distance function d –Irrelevant data needs to be removed –Slower than model-based recommendations	–Recommend similar peers, or contents to learners –Cluster learners in groups
8. Vector-based models (TF-IDF, Singular value decomposition, Matrix Factorisation)	Vector-based approaches characterise items and users as vectors of factors in a 3D space. A high correlation between an item and a user can be used as recommendation but also predictions can be created.	–Suitable for sparse datasets –Can take temporal differences into account –Can take various implicit information into account does not need explicit ratings	–Content depended (Items with same context but different terms are not matched) –User keywords have to match semantic space	–Useful to monitor and predict learner performance –Can adapt to increased knowledge level of learners –Can mark learning resources that are not popular anymore

 

Framework for the analysis of Recommender Systems