Knowledge based RS

For example, apartments are not bought very frequently which makes it rather infeasible to collect numerous ratings for one specific item. Furthermore, users of recommender applications would not be satisfied with recommendations based on years-old item preferences.

Knowledge-based recommender technologies help to tackle these challenges by exploiting explicit user requirements and deep knowledge about the underlying product domain.

They have no cold-start problems since requirements are directly elicited within a recommendation session, but they do suffer from the so-called knowledge acquisition bottleneck in the sense that knowledge engineers must work hard to convert the knowledge possessed by domain experts into formal, executable representations.

Limitations

• Cost of knowledge acquisition
  • From domain experts (determination of the recommendation rules/ constraints)
  • From users (determination of user requirements)
• Accuracy of preference models
  • Very fine granular preference models require many interactions cycles
• Independence assumption can be challenged
  • Preferences are not always independent of each other

There are 2 types of Knowledge-Based RS:

• Constraint-Based RS
• Case-Based RS

Constraint-Based RS

Constraint-based recommenders predominantly exploit predefined recommender knowledge bases that contain explicit rules about how to relate customer requirements with item features.

Constraint Satisfaction Problem (CSP)

Constrain satisfaction problem can be described by a tuple $(V,D,C)$

• $V$ is a set of variables
  • Consumer Requirements, $V_C$
  • Item Features, $V_{PROD}$
  • $V=(V_C\cup V_{PROD})$
• $D$ is a set of finite domains for $V$
• $C$ is a set of constraints that describe the combination of values and variables that can simultaneously take
  • Filter, $C_F$: IF the user requires A THEN the item should possess feature B
  • Compatibility, $C_R$: IF user requires A1 THEN she should propose A2>200
  • Product, $C_{PROD}$: defines available product assortments and features
• Customers query (REQ) encoded as a constraint over the variable $V_C$ and $V_{PROD}$

Recommendation task: identifying a set of products matching (RES) based on user requirements (REQ)

• Find a set of user requirements such that a subset of items fulfils all constraints
  • Ask the user which requirements should be relaxed/modified such that some items exist that do not violate any constraint
• Find a subset of items that satisfy the maximum set of (weighted) constraints considering REQ
  • All proposed items have to fulfil the same set of constraints
• Rank items according to the weight of satisfying constraints, based on REQ
  • Rank items based on the ratio of fulfilled constraints
• Conjunctive query is an integration of
  • Filter condition $C_F$
  • User Requirement REQ

Defaults in CSP

Users might be unsure about what options to select for the item they want or simply do not know the technical details

• Static: One default value is specified for each user property
• Dependent: Defined on different combinations of user requirements
• Derived: Exploit previous interactions to derive default values
  • 1-nearest neighbour
    • Predicts the default value for 1 or sets of user requirements
    • Selects the entry that is most similar to the current user requirement
  • Weighted majority voter
    • Propose default value for user property based on the voting of a set of similar items for that specific property.

Dealing with unsatisfied requirements

• Rank the item based on Multi-Attribute Utility Theory (MAUT)
  • Each item is evaluated according to a predefined set of dimensions that provide a general perception of the basic item properties.
  • Ex: quality and economy are dimensions in the domain of digital cameras. If $price \le 250$, $quality=5$, $economy=10$ $$utiliy(p)=\sum_{j=1}^{n} {interest(j)*contribution(p,j)}$$

Case-Based RS

Case-based recommenders determine recommendations on the basis of similarity metrics. Items are retrieved using similarity measures that describe to which extent item properties match user requirements.

• Early versions (query-based approach)
  • Customers have to specify their requirements and the system retrieves items similar to user requirements.
  • Drawback: customers maybe don’t know what they are seeking
  • Solution: development of browsing based approach
    • Users who do not know what they are seeking, are navigating in the item with the goal to find useful alternatives
• State-of-the-art (critique-based recommender)
  • Integrates query-based approach with the browsing-based approach
  • Supports identification of the most similar items
  • Allows users to change preferences without being forced to specify the exact value for item attributes

Distance Similarity Measures

$$similarity(p,REQ)=\frac {\sum_{r\in REQ} {w_r * sim(p,r)}} {\sum_{r\in REQ} w_r }$$

• $sim(p,r)$ expresses the distance of r from each item attribute value $\phi_r(p)$
• $w_r$ is the importance weight for requirement $r$
• Maximize certain properties, i.e., resolution of the camera, more is better (MIB) $$sim(p,r)=\frac{\phi_r(p)-\min(r)}{\max(r)-\min(r)}$$
• Minimize certain properties, i.e., price of the camera, less is better (LIB) $$sim(p,r)=\frac{\max(r)-\phi_r(p)}{\max(r)-\min(r)}$$
• There are situations in which the similarity should be based solely on the distance to the originally defined requirements, i.e., certain run time or monitor size $$sim(p,r)=1-\frac{\left | \phi_r(p)-r \right |}{\max(r)-\min(r)}$$

Interacting with Case-based recommenders

• Critiquing is an effective way to support such navigations (browsing)
• Customers specify their change requests that are not satisfied by the current item
• Goal is to achieve time savings in the item selection process as well as achieve the same recommendation quality as query-based approaches
• Operating over multiple properties can improve the efficiency of recommendation dialogues (Ex: cheaper and more pixels)
  • Reduce the number of critiquing cycles
  • Faster progress into item space
• Item recommendation is based on
  • Similarity degree of new recommended item $(c_i)$ with user requirement (REQ)
  • Compatibility degree of recommended item $(c_i)$ with already selected compound critique $(CC_U)$ $$compatibility(c_i,CC_U )=\frac {\left | {cc_i\in CC_U:satisfies(cc_i,c_i )} \right |} {\left | CC_U \right |}$$
  • Quality measure for the recommended item $(c_i)$ $$quality(c_i, REQ, CC_U)=compatibility(c_i, CC_U )*similarity(c_i, REQ)$$