2016 TKDE Prefix-Adaptive and Time-Sensitive Personalized Query Auto Completion

**Main Problem:**
This paper proposed time-sensitive and personalized query auto completion (QAC), named hybrid QAC. They handle the long-tail prefixes. Given a threshold N, we define a prefix p to be a long-tail prefix if the number of returned query completions is smaller than N. In other words they defined long-tail prefixes based on the number of returned query completions rather than the number of characters in the prefix.
Time-sensitive: the recent and long-term query frequencies are considered to predict future popularity.
Personalized: This model uses the combination of two similarity scores: recent queries in the current session and those of the same user issued before. (caters for individual users)

**Input-Output:**

Input: input prefix p by user + calculating Time-sensitive and Personalized features
Output: a ranked list of queries for prefix p

**Previous Works and their Gaps:**
- To forecast the query frequencies, Radinsky[] proposed a long term time series modeling approach
- Cai and de Rijke[] proposed a learning-based QAC approach. Their model derived features from similar queries and semantically related terms
- considered profiles to extract user-based features to model the likelihood
that a user will issue certain queries.

Gap: In previous works, query completions are computed globally and for a given prefix, and all users are presented with the same list of completions, and none of them deal with long-tail prefixes

Result:
They use two data sets: AOL and SnV (the largest audiovisual archives in Europe).
The proposed method improves the Mean Reciprocal Rank (MRR) scores between 4 and 8 percent on a web search log and on a query log from an audiovisual archive, compares to the state-of-the-art time-sensitive QAC baseline

Gap of this work:
1-In order to personalized the model they just considered users' search log
Code:
Not available