2022 WSDM Improving Personalized Search with Dual-Feedback Network

**Main Problem:**
This paper proposed a model to improve personalized search. They indicate that queries have multiple intentions, one of them is the positive intention which refers to users' true search intentions. The other is a negative intention which refers to what users are not interested in a search action. Their model learns these two types of intentions in a search log based on the given query. Their model learns positive and negative intention in both long-term search and short-term search. Clicking documents consider a positive intention and skipping documents represents negative intention. They used these intentions in building users' profiles as a personalization search.

**Input-Output:**

Input: long-term positive and negative intention + short term positive and negative intention based on the given query
Output: rerank documents

**Previous Works and their Gaps:**

The previous works mostly consider the positive intention of the user to understand the user search intent. This paper claimed that considering both intentions could improve the personal search

**Result:**
They compared their model to other personalized models. This model improved personalized search performance by 1.15% based on the MAP metric.
They also compared their model with the ad-hoc models to show the personalized model performs based on deep learning are better than the ad-hoc model. Ad-hoc search models consider the relevance between queries and documents in ranking documents.

**Data Set:**
They conducted experiments on the Aol dataset and commercial dataset (they didn't name it)

Gap of this work:
In order to capture positive and negative users' search intentions to build users' profiles, users' comments analysis on social networks can be considered to improve the users' profile in personalized search. For example, negative comments can be categorized as negative intentions and positive comments can be categorized as positive intentions. So, we have more complementary information about users' interests and users' search intent would be captured better.

Code: