2020 SIGIR Employing Personal Word Embeddings for Personalized Search

[ZahraTaherikhonakdar]

Main Problem:

his paper proposed a new model for web search personalization. They develop a personalized word embedding model based on users' search logs to obtain each user intent. Their aim is to create different word representations of the same word for different users. In this case, they can reduce the ambiguity of words with multiple meanings. For example, in their word embedding model, "Apple " would be represented differently for each user. For useri, who is an IT engineer, the representation is "Apple+user i", the representation for user j who is a farmer would be "Apple+user j". They are aware of changing user intent over time, so they develop an approach to obtain the user's intent and train the model based on new data.

Input-Output:

phase: Input: users' search log --> representation of words as vectors--> query reformulation Output: ranking documents based on each user interest

Previous Works and their Gaps: They divided personalized models in web search into two categories: 1- traditional personalized search model --> consider users' click information and issued query. 2- deep learning-based Personalized search--> train a model on users' query or search log and build a user profile Unlike previous work, this paper looks into personalization from a different perspective with aim of eliminating word ambiguity based on each user without a user profile and obtaining the user interest over time.

Result: They compare their proposed model with other personalized models (User profile-based personalized/Embedding based personalized search model) and based on the results they achieve significant improvement in MAP and MRR metrics.

Data Set: They used the AOL data set and commercial dataset(they didn't indicate the name).

Gap of this work: Based on the results, the only point that comes to my mind is, in the updating stage for obtaining users' future interest we could consider users' social information too. Also, this could be helpful for a better understanding of each user's interest in general and reducing users' intention ambiguity.

Code:

[hosseinfani]

@ZahraTaherikhonakdar thank you. interesting idea => personalized word embedding