Topic-based personailization
\begin{equation}
\beta \Pr (d│q)+(1−\beta) \Pr_{cat} (d│q, u)
\end{equation}
\begin{equation}
\Pr_{cat} (d|q, u)= \sum_c \Pr (c│d) \Pr (c|u)
\end{equation}
- Metric: MRR - improved a lot on the ambiguous queries
- Main idea:
- consider the user’s and doc’s topic-category matches
- A personalized SE consider
- Relevance
- The query-independent value of scores
- The topic is interested in or not
Long-short term personalization
10 features for atypical discovery
- Query length
- Query length divergence
- SAT Reading Level
- SAT Reading Level Divergence
- Topic divergence
- Ratio of noun
- Verb ratio divergence
- Adjective ratio divergence
- Longest query position
- Question ratio
Divergence = distance of current session to historical vocabs/topic categories/features