Main Problem:
This paper proposed a graph network in a recommender system with the aim of understanding users' intent. They claimed that paying attention to two types of data would boost the performance of the recommender system. First, they consider different weights for a user's different behaviors based on the target behavior. For example, in online shopping, the target behavior is shopping, and add-to card or suggest to friends is auxiliary behaviors that this paper considered for predicting users' intent and put different importance on them based on the target behavior (here is a suggested list of items for shopping based on users behaviors). They showed this information as user-to-item relation in the model. Second, they consider the semantics of different types of behaviors. In other words, they detect the relevance between two items that a user clicked and viewed. This relevance could have an influence on the user's behavior (here buy or not). For example, if the users clicked and viewed iphon and airpod, these two items may have some relatedness that affects the users' behavior. They showed this kind of information as an item-to-item behavior in the model (this information solved the cold-start problem that previous models have. cold-start refers to a situation when there is not enough information to predict, for example, a user's shopping behavior).
Input-Output:
Input: the user-item and item-item information based on the target behavior
Output: a recommendation model that predicts the probability of a user interacting with items
Previous Works and their Gaps:
The previous works in the recommendation system only considered one key behavior. For example, for suggesting items for a user to buy, they only consider a user's purchase actions. Recommender systems that consider multi-behavior in their models suffer from shortcomings. First, they did not consider the importance of different behaviors to predict the target model. Second, they did not pay attention to users' co-behaviors. (item-item information which I explained before)
Result:
They compared their model to single-behavior models and multi-behavior models. They conducted experiments on two datasets, namely Tmall and Beibei. They improved the performance by 39.28% on Recall, and 14% on NDCG compared Tmall dataset and 14.65% and 11.07% on the Beibei dataset.
Data Set:
They conducted experiments on two datasets namely Tmall and Beibei. They include information about users' purchasing, carting, and clicking. The Tmall dataset is publicly available at: https://tianchi.aliyun.com/dataset/dataDetail?dataId=649
Gap of this work:
This work could predict users' intent better than the previous works. One thing that could enhance the performance of this model is considering users' search logs or likes and comments on online shops on social networks
Main Problem: This paper proposed a graph network in a recommender system with the aim of understanding users' intent. They claimed that paying attention to two types of data would boost the performance of the recommender system. First, they consider different weights for a user's different behaviors based on the target behavior. For example, in online shopping, the target behavior is shopping, and add-to card or suggest to friends is auxiliary behaviors that this paper considered for predicting users' intent and put different importance on them based on the target behavior (here is a suggested list of items for shopping based on users behaviors). They showed this information as user-to-item relation in the model. Second, they consider the semantics of different types of behaviors. In other words, they detect the relevance between two items that a user clicked and viewed. This relevance could have an influence on the user's behavior (here buy or not). For example, if the users clicked and viewed iphon and airpod, these two items may have some relatedness that affects the users' behavior. They showed this kind of information as an item-to-item behavior in the model (this information solved the cold-start problem that previous models have. cold-start refers to a situation when there is not enough information to predict, for example, a user's shopping behavior).
Input-Output:
Input: the user-item and item-item information based on the target behavior Output: a recommendation model that predicts the probability of a user interacting with items
Previous Works and their Gaps:
The previous works in the recommendation system only considered one key behavior. For example, for suggesting items for a user to buy, they only consider a user's purchase actions. Recommender systems that consider multi-behavior in their models suffer from shortcomings. First, they did not consider the importance of different behaviors to predict the target model. Second, they did not pay attention to users' co-behaviors. (item-item information which I explained before)
Result: They compared their model to single-behavior models and multi-behavior models. They conducted experiments on two datasets, namely Tmall and Beibei. They improved the performance by 39.28% on Recall, and 14% on NDCG compared Tmall dataset and 14.65% and 11.07% on the Beibei dataset.
Data Set: They conducted experiments on two datasets namely Tmall and Beibei. They include information about users' purchasing, carting, and clicking. The Tmall dataset is publicly available at: https://tianchi.aliyun.com/dataset/dataDetail?dataId=649
Gap of this work: This work could predict users' intent better than the previous works. One thing that could enhance the performance of this model is considering users' search logs or likes and comments on online shops on social networks
Code: