Matching Recommendation Technologies and Domains - Part 1

 

 

This paper is written by Burke and Ramezani, they introduce a way of mapping between domain characteristics and recommendation technologies, which is very interesting point to talk about.

Paper Info.:
Burke, R. & Ramezani, M. Matching Recommendation Technologies and Domains. In Kantor, P., Ricci, F., Rokach, L. & Shapira, B. eds. Handbook of Recommender Systems. Springer, 2010. (PDF) 

 

Paper Abstract:

Recommender systems form an extremely diverse body of technologies and approaches. The chapter aims to assist researchers and developers identify the recommendation technology that are most likely to be applicable to different domains of recommendation. Unlike other taxonomies of recommender systems, our approach is centered on the question of knowledge: what knowledge does a recommender system need in order to function, and where does that knowledge come from? Different recommendation domains (books vs condominiums, for example) provide different opportunities for the gathering and application of knowledge. These considerations give rise to a mapping between domain characteristics and recommendation technologies.

 

What is a Recommender system?

 

A recommender system is defined by a particular kind of semantics of interaction with the user: “any 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”. This expansive definition makes the scope of recommender systems research quite broad, but it fails to give much guidance to the implementer.

 

Related Work (Recommender Systems Taxonomies)

 

The Author introduced 5 papers, which offered different taxonomies.

1. R. Burke, “Hybrid recommender systems: Survey and experiments,” User Modeling and User- Adapted Interaction, vol. 12, no. 4, pp. 331–370, 2002.
Contribution: Five recommendation techniques, Advantages and Disadvantages per each, and hybrid methods.
 
2. P. Resnick and H. R. Varian, “Recommender systems,” Commun. ACM, vol. 40, no. 3, pp. 56– 58, 1997. 
Contribution: Identifies 5 dimensions, which characterize properties of users' interactions with the recommender and the aggregation methods of users' evaluations (ratings).
 
3. J. B. Schafer, J. A. Konstan, and J. Riedl, “E-commerce recommendation applications,” Data Mining and Knowledge Discovery, vol. 5, no. 1-2, pp. 115–153, 2001.
Contribution: Taxonomy of collaboratice e-commerce recommender applications into 3 categories: functional I/O, the recommendation method, and other design issues.
 
4. M. Montaner, B. L´opez, and J. L. D. L. Rosa, “A taxonomy of recommender agents on the Internet,” Artif. Intell. Rev., vol. 19, no. 4, pp. 285–330, 2003.
Contribution: using two main criteria: user profile generation and maintenance, and user profile exploitation (Common patterns are extracted by an analysis of systems in the same domain).
 
5. G. Adomavicius and A. Tuzhilin, “Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions,” IEEE Transactions on Knowledge and Data Engineering, vol. 17, no. 6, pp. 734–749, 2005.
Contribution: Comparing between 3 main categories (content, collaborative, and hybrid.. omitting knowledge-based and classifies recommenders in each category into either heuristics-based or model-based).

Paper Contribution

AI-centric approach, focused on the knowledge sources required for recommendation and the constraints related to them. The chapter discusses the applicability of different recommendation techniques to different types of problems and aims to guide decision making in choosing among these techniques. As such, it might be considered to serve as a sort of recommender for recommender system implementers.

Knowledge Sources

 

• Social: Knowledge about the larger community of users other than the target user.
• Individual: Knowledge about the target user.
• Content: Knowledge about the items being recommended and, more generally, about their uses.

The Recommender System use the knowledge sources based on the type of problem it aimed to solve.

Recommendation types vs Knowledge Sources

 

• Collaborative recommendation matches an individual knowledge source with a social knowledge source of the same type and extrapolates the target user’s preferences from his or her peers. Usually, in collaborative recommendation, individual requirements are not used, or applied very simply as filters.


• Content-based recommendation on the other hand is individually-focused, using item features and user opinions to learn a classifier that can predict user preferences on new items.


• Knowledge-based recommendation is more of a catch-all category in which the recommender applies any kind of domain knowledge more substantive than item features.

Domain Characteristics

A domain of recommendation is the set of items that the recommender will operate over, but may also include the set of aims or purposes that the recommender is intended to support. These characteristics of the domain affect the availability and utility of different knowledge sources.

Six important characteristics of the domain that an implementer should consider:

1. Heterogeneity: A heterogeneous item space encompasses many items with different characteristics and most importantly, different goals they can satisfy.

2. Risk: Recommendation domains can be distinguished by the degree of risk that a user incurs in accepting a recommendation.

3. Churn: Recommender systems are used in domains with long-lived items like books, but they are also used in domains where the value or relevance of an item has a very short time span, such as news stories. A high churn domain is one in which items come and go rapidly.

4. Interaction Style: In systems in which the user makes no special effort to interact with the recommender system, the system extracts the implicit expressed preferences from user behaviour.

5. Preference Stability: User preferences can also have varying degrees of duration.

6. Scrutability: Certain applications (for example, high-risk ones) may require that the system be able to explain its recommendations, to answer questions like “why was this item Matching Recommendation Technologies and Domains recommended?” Such explanations enhance user confidence that a recommendation is appropriate and increase the liklihood of recommendations being accepted.
 

Hybrid Recommender Systems

 

 

 

 

This is a Review for a paper written by Robin Burke, I'll try to summarize it briefly and mention the most important points on this paper.

 

Paper Info.:

Burke, R. Hybrid Recommender Systems: Survey and Experiments. User Modeling and User-Adapted Interaction. 12(4), pages 331-370. [PDF]

 

Paper Abstract:

Recommender systems represent user preferences for the purpose of suggesting items to purchase or examine. They have become fundamental applications in electronic commerce and information access, providing suggestions that effectively prune large information spaces so that users are directed toward those items that best meet their needs and preferences. A variety of techniques have been proposed for performing recommendation, including content-based, collaborative, knowledge-based and other techniques. To improve performance, these methods have sometimes been combined in hybrid recommenders. This paper surveys the landscape of actual and possible hybrid recommenders, and introduces a novel hybrid, EntreeC, a system that combines knowledge-based recommendation and collaborative filtering to recommend restaurants. Further, we show that semantic ratings obtained from the knowledge-based part of the system enhance the effectiveness of collaborative filtering.
 

Recommender System Definition Past vs. Present

 

Recommender systems were originally defined as “people provide recommendations as inputs, which the system then aggregates and directs to appropriate recipients”, like using experts knowledge as input for the system to enrich its ability to recommend to people according to the given knowledge.

The term now has a broader connotation, describing any 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.

 

Recommender System vs. Search Engines (Information Retrieval)

 

It is the criteria of “individualized” and “interesting and useful” that separate the recommender system from information retrieval systems or search engines. The semantics of a search engine are “matching”: the system is supposed to return all those items that match the query ranked by degree of match. Unlike recommender system it aims for personalizing the content based on assumptions (Recommender system approaches) which find something similar to user's taste and desire.

 

Recommender System Hybridization

 

All of the known recommendation techniques have strengths and weaknesses, and many researchers have chosen to combine techniques in different ways.

 

Recommendation Techniques

 

 

Comparing Recommendation Techniques 

 

 

Hybrid Recommender Systems Methods

 

 

Possible Recommendation Hybrids

 

The author explains his "EntreeC" experiment by combining Knowledge-based and Collaborative Filtering Approach using Cascade hybrid method and also explained how did he convert his semantic ratings into numeric.

 

The author presented an important information about the Order of Implicit Ratings strength (from Nichols paper,1997).

 

 

 

Hybrid or No Hybrid Choice

 

While the space remains to be fully explored, research has provided some insight into the question of which hybrid to employ in particular situations. The hybridization strategy must be a function of the characteristics of the recommenders being combined.

 

With demographic, content and collaborative recommenders, this is largely a function of the quality and quantity of data available for learning. With knowledge-based recommenders, it is a function of the available knowledge base.

 

We can distinguish two cases: the uniform case, in which one recommender has better accuracy than another over the whole space of recommendation, and the non-uniform case, in which the two recommenders have different strengths in different parts of the space.

 

If the recommenders are uniformly unequal, it may make sense to employ a hybrid in which the inaccuracies of the weaker recommender can be contained: for example, a cascade scheme with the stronger recommender given higher priority, an augmentation hybrid in which the weaker recommender acts as a “bot” contributing a small amount of information, or a meta-level combination in which the stronger technique produces a dense representation that strengthens the performance of the weaker one.

 

In the non-uniform case, the system will need to be able to employ both recommenders at different times. A switching hybrid is a natural choice here, but it requires that the system be able to detect when one recommender should be preferred. Feature combination and mixed hybrids can be used to allow output from both recommenders without having to implement a switching criterion. More research is needed to establish the tradeoffs between these hybridization options.