2020 SIGIR Personalized Query Suggestions

**Main Problem:**
They proposed a sequence-to-sequence-model–based query suggestion framework that is able to model structured, personalized features and unstructured query texts naturally. For personalization, they use Linkedin members' profiles.

**Input-Output:**

Input: query suggestion request + user feature
Output: query suggestion response

**Previous Works and their Gaps:**
- Mei [] proposed a bipartite graph– based method to derive query suggestions.
Their method achieved personalization by attaching a unique user
identifier to each query to form a new “pseudo query”.
- Jayanthi [8] used phrase semantics instead of term semantics to derive personalized query suggestions.
- Sordoni [], the authors explored the use of hierarchical recurrent encoder-decoder for generating query suggestions. The methodology is the most similar to this paper. However, this paper has very different focuses: Sordoni [] were mostly interested in the context-awareness of the query suggestions, where context is defined by past search history, while this paper focuses are mainly on the personalization with user features and the approach is also tested in a large scale industrial setting (Linkedin).

**Result:**
Their model improved query suggestions by 5.62% and the overall search results by 0.3%.
They divided Linkedin members into two categories: passive and active job seekers based on their recent job-related activities. They realize that personalization has a more significant impact on passive job seekers because those suggestions are relevant to their background and intent.

**Data Set**
They mined English search query logs at LinkedIn.

Gap of this work:
1- Although users' profiles are considered, users' activities and time are not considered.

Code:
Not available