Recommender Agent

Definition

Recommender Agent

Relevance Score estimation

A recommender agent is a software agent that uses information about a user and a set of available items to estimate which items are relevant to that user and recommend them accordingly.

Given a user model $u$, a set of items $I$, and optionally a context $c$, the agent may estimate a relevance score

$$r(u,i,c)$$

for each $i\in I$ and use these scores to rank or select items for recommendation.

Xiao & Benbasat, MISQ, 2007

Recommendation agents (RAs) are software that elicits the interests or preferences of individual consumers for products, either explicitly or implicitly, and make recommendations accordingly. RAs have the potential to support and improve the quality of the decisions consumers make when searching for and selecting products online. ¹

Paradigms

Recommendation paradigms differ mainly in the evidence used to estimate the relevance of an item $i$ to a user $u$. A score may depend on

$$r(u,i,c)=F\left(p_u,x_i,A,K,c\right),$$

where $p_u$ is the user’s profile, $x_i$ describes the item, $A$ records interactions between users and items, $K$ contains domain knowledge, and $c$ is the current context. Each paradigm relies on a different subset of this information.

User-profile-based

User-profile-based recommendation uses attributes or preferences recorded directly in the user’s profile, such as age, location, stated interests, or selected requirements. It recommends items whose target audience or attributes match that profile.

The profile may be elicited explicitly or inferred from behaviour. This is a broad personalisation strategy and can supply information to the more specific paradigms below.

Collaborative

Collaborative recommendation uses the interaction matrix $A$ rather than requiring a semantic description of each item. It estimates $r(u,i)$ from ratings, purchases, clicks, or other behaviour of users with similar interaction patterns, or from items consumed by similar groups of users.

Its guiding question is: Which items were valued by users whose behaviour resembles this user’s behaviour?

Content-based

Content-based recommendation compares item features $x_i$ with a preference profile learned from items the user previously valued. For example, a user who repeatedly watches science-fiction films may receive other films with similar genres, topics, or creators.

Its guiding question is: Which items resemble the items this user has valued before?

Knowledge-based

Knowledge-based recommendation uses domain knowledge $K$ to connect the user’s stated needs and constraints to item properties. It may apply rules, constraints, or a utility function, such as requiring a flat to satisfy a maximum price, minimum size, and acceptable location.

Its guiding question is: Which items satisfy this user’s present requirements? This is particularly useful when interactions are sparse or purchases are too infrequent for a behavioural history to be informative.

Hybrid

A hybrid recommendation system combines two or more paradigms. For example, it may use collaborative evidence when sufficient interaction data exist and content or knowledge-based evidence for new users and items.

Footnotes

1. E-Commerce Product Recommendation Agents - Use, Characteristics, and Impact ↩