2019 WWW Click Feedback-Aware Query Recommendation Using Adversarial Examples

**Main Problem:**
A novel neural network based method, Click Feedback-Aware Network (CFAN), was proposed to provide feedback-aware query suggestions. To capture the underlying search intent of users CFAN considers: 1- user clicks on previous suggested queries (user feedback). 2- the issued search query sequence

**Input-Output:**
Phase 1: Candidate Generation:
Input : a search query Q
Output : set of suggestion candidates queries

Phase 2: Input : set of suggestion candidates queries
Output : ranked list of query suggestions

**Previous Works and their Gaps:**
1- employs a hierarchical encoder-decoder model to achieve context-aware query suggestion
2- introduces the copy mechanism in addition to employing the encoder-decoder framework to generate suggested queries.
3- employs a feedback memory network to model titles of user clicked URLs
4- Reformulation inference network (RIN) is proposed explicitly incorporates query reformulations during training.
Gap: In none of the above neural network-based works previously user clicked suggestions aren’t explicitly incorporated as feedback

**Result:**
They use a search log collected from one of the largest search engines in the world. (did not mention the name of the data set). They used 3,684,008 training queries within 493,864 sessions and 406,063 test queries within 55,141 sessions. They employed Mean Reciprocal Rank (MRR) score as the evaluation metric. The performance of CFAN id evaluated against the following state-of-the-art methods:
Query-based Variable Markov Model (QVMM), Feature-Based Suggestion (FBS), Reformulation-based Completion (RC), Most Popular Suggestion (MPS), Hybrid Suggestion (Hybrid), Hierarchical Recurrent Encoder-Decoder (HRED), Reformulation inference network (RIN)

**Gap of this work:**
1- considering users previous clicks in one session to suggest a query isn’t a proper method because a user could simply change the search intent in one session.
2- To capture the user’s search intent the temporal and personal information are not considered in this paper.

**Code:**
Not available