How Different Features Contribute to the Session Search?

Learning to Personalize for Web Search Sessions

“…Our approach extends beyond similar approaches. Previous work in this category uses features from the current session only [13,42] or related sessions from the query logs submitted by the same user [2] or other users [22]. Our results show that our approach significantly outperforms existing session search approaches on four of session search's test collections.…”

“…Sessions belonging to the same cluster are considered related. Li et al [22] investigated the effectiveness of four classes of features with respect to: current query; query change; whole session; and related sessions. To determine topically related sessions, they estimate an LDA topic model over clicked documents in the entire query logs.…”

“…In essence, this task is an online task that needs to be performed at query time. The method proposed by Li et al [22] requires an LDA topic model to be estimated over all clicked documents which is prohibitive in practice. Also, these clustering methods require the number of topics to be set a priori.…”

“…This pruning rule presupposes that for two sessions to be related, they must be relevant to at least one common social position. This captures the semantic similarity between two candidate sessions without the need to run a computationally expensive topic modelling algorithm on the entire clicked documents as in Li et al [22]. Features 8, 9 and 10 rely on identifying each session's social positions.…”

“…LTR-LDA uses the related sessions' features in table 4 in addition to all the general features in table 3. Related sessions were identified using the LDA-based approach that was discussed in section 3.4 [22]. These systems are compared with LTR-SP which uses the full set of features.…”