Making Personalised Flight Recommendations using Implicit Feedback
As e-commerce has become more popular, the problem of information overload has come to the fore. Recommender systems that reduce the information overload problem are becoming more common. However, the problem with many recommender systems is that they are associated with a high cost of learning customer preferences (in terms of cognitive load). We describe the Personal Travel Assistant (PTA), a flight recommender application that uses case-based reasoning (CBR) to overcome these problems. The PTA allows users to search multiple flights providers concurrently and recommends flights based on their individual travel preferences. These preferences are implicitly learned from observations of user behaviour. When the user purchases a flight, the PTA uses the selection of a preferred flight to discover and refine the user's overall travel preferences. These preferences are stored in a user-model as sets of cases representing their interactions, which are used to provide personalised recommendations. The PTA makes recommendations taking into account the context in which the flights were offered. It uses features from the request to determine this context, e.g. the duration of the trip. We perform evaluations of contextual recommendations that support our view that user preferences change depending on the context of the session. We further improve recommendation accuracy by storing and personalising similarity measures in the user-model. The PTA alters the relative importance of features in the personal similarity measure based on implicit user feedback, e.g. increasing the importance of price at the cost of stop-over time in a multiple hop flight. We also investigate cooperative components to extend our recommendation strategies. These allow users to reuse the information learned from other users when they encounter new situations. However, these techniques are not as successful as we had hoped. We discuss these components in relation to other work on collaborative recommendation and suggest that the standard approach is unsuited to the PTA's context-based recommendation strategy. The strength of CBR in the e-commerce domain stems from its reuse of the knowledge base associated with a particular application. Since case data may be one aspect of a company's entire knowledge system, it is important to integrate case data easily within a company's IT infrastructure, providing in effect a case-based view on relevant portions of the company knowledge base. We describe CBML, an XML-based Case Mark-Up Language we have developed to facilitate such integration.