2.9 Recommender Systems
2.9.4 Case-based recommender systems
The customers who purchase complex products such as financial services, laptops, or digital cameras need both information and intelligent interaction mechanisms that support the selection of appropriate solutions. One particular class of knowledge- based RSs is case-based recommenders (Smyth 2007, Bridge et al. 2005). This style of recommenders is suitable to help customers identify complex products. Indeed, for products like laptops, digital cameras, cars or houses, or even financial services, customers are not very keen on depending totally on other customers’ reviews to be comfortable with a recommendation and make the decision to purchase products. For products that are relatively expensive or professionally critical, people are willing to put enough effort for the purpose of reaching a choice which satisfies the user’s needs as best as possible with a minimum of risk (e.g., without substantial financial loss). In fact, people can spend time in interacting with the system in case they want it to recommend a suitable product which has a great value for them (e.g., a house).
Accordingly, such products are usually constrained by a set of features (for ex- ample, the digital camera has features like price, optimal zoom, resolution, etc.) for
2.9. Recommender Systems
which customers are able to specify particular preferences (e.g., to specify values, etc.) to filter out the available large data set.
Such organized information about products allow case-based recommenders to adopt and provide required formalisms for intelligent interaction mechanisms that support the selection of the most appropriate solutions within a sizeable set of pos- sible items.
Instead of relying on the collective preferences that collaborative filtering uses, case-based recommendation suggests explicit sales dialogues that emulate the con- versation between a customer and a sales assistant. These dialogues help to proac- tively suggest products to a given customer by predicting her preferences on prod- ucts.
2.9.4.1 Single-shot recommender systems
Case-based recommendation systems used to operate in a “single-shot” fashion. When a “single-shot” RS receives a query from the user, it recommends a set of products to the user only once. If the user is not satisfied, there is only one way to proceed in order to readjust the “shot”: to amend the query and start again. In “single-shot” RSs, the same results are often returned to the user despite being of no interest to the user (Smyth 2007).
This recommendation model supposes the user knows what she is looking for to a point that the system does not consider to have more than a single “exchange” with the user. Figure 4.1 in Chapter 4 illustrates the single-shot recommendation scenario. This would not be supported by the actual extreme diverse products that are available in the web. In fact, the ability to refine the search in large catalogs through multiple interactions could be helpful for the users. Using multiple interactions, a user can navigate a product space much more effectively while taking the time to express her preferences progressively.
2.9.4.2 Conversational recommender systems
Generally speaking people do not state their preferences up-front because initially they only have a vague idea of the product they would like to have (Bridge et al. 2005). Usually, criteria about the product the customer would like to purchase are specified during the dialog with the seller. This is still the case even for knowledgable customers in the domains where expert users need to be assisted because available products dynamically change. A distinctive example is the list of special offers (e.g., flight tickets) which change frequently.
In order to hold natural-like and interactive dialogue, conversational RSs (Bridge et al. 2005) were proposed. These systems support a dialogue where, at each stage,
the system can select one from a set of available system actions, e.g., recommend some products or ask the user for more information. The particular action selected by the system is determined by its recommendation strategy.
The recommendation strategy usually characterizes conversational RSs since it determines the action plan adopted by the recommender while interacting with the user, in order to help her reach her ultimate goal: obtain the most suitable possible offer from the system (e.g., product, service or combination of both). The recommen- dation strategy ultimately depends on the complexity of the user’s goal, as it should accompany and guide the user while she is searching for her target. For instance, in the prototype of the Austrian Tourism portal (Mahmood & Ricci 2007, Mahmood et al. 2008), there are three possible targets for the user: 1) “Build a Travel Plan ”, 2) “Window Shop ”, i.e., just browse through the products, and 3) “Book a Product ”, i.e., search for a single product and add it to the travel plan.
In a conversational scenario, which is depicted in Figure 4.1 in Chapter 4, the strategy is implemented by specifying the particular actions that the system executes at each stage of a given interaction session. Indeed, at each interaction cycle, the system can either request from the user a preference or propose a product to the user. The user can either answer the question posed or criticize the system proposal. As the recommendation strategy offers various ways of interacting with the user, conversational RSs are able to recommend a variety of products for diverse categories of users.
Requirements acquisition is regarded as a key element in e-commerce. The prod- uct search, which probably needs a filter-based retrieval, can take place in tandem with preference elicitation. Bridge et al. (Bridge et al. 2005) described two ways of requirements elicitation by which conversational recommenders can be identi- fied: the first approach is called navigation-by-asking according to which the RS sets a dialogue with the user during which the RS selects and asks questions to the user, whether up-front (e.g., form-filling) or incrementally. The user’s prefer- ences are elicited when the user answers the questions (Goker & Thompson 2000, Doyle & Cunningham 2000, Shimazu 2001, Shimazu 2002, Schmitt 2002b, Thomp- son et al. 2004a). The second approach is called navigation-by-proposing according to which the RS may show the user intermediary products so that it gets a form of feedback from the user’s critiques to the proposed products when invited to give feedback about what was presented to her (Burke et al. 1996, Burke et al. 1997a, Shimazu 2002, Faltings, Pu, Torrens & Viappiani 2004, McCarthy et al. 2004, Pu & Faltings 2004, McCarthy, McGinty, Smyth & Reilly 2005, McCarthy, Reilly, McGinty & Smyth 2005, Chen & Pu 2009).
Conversational RSs may implement only the asking/answering conversation mode (Linden et al. 1997), only the proposing/ criticizing conversation mode (Burke 2002),