WiMi Hologram Cloud Inc., a leading global Hologram Augmented Reality Technology provider, announced that a recommendation model based on heterogeneous information network (HIN) has been developed, bringing a breakthrough in the personalized recommendation. The recommendation model based on HIN, which consists of different types of nodes and multiple types of relationships, can better portray complex relationships in the real world.
The recommendation model based on HIN aims to solve the problems of current Internet recommendation models, including data sparsity, misleading information extraction, and loss of recommended useful information. These problems are challenging for traditional recommendation models, and WiMi’s HIN-based recommendation model can solve these problems.
Data sparsity is a common problem nowadays, especially in the case of limited user behavior data. Traditional recommendation models such as collaborative filtering have difficulty in accurately capturing user interests and preferences. This model can alleviate the problem of data sparsity by utilizing multiple meta-paths to describe the relationship between users and items, which can be used to alleviate the data sparsity problem by transferring information across meta-paths. Even if there is a lack of user-item interaction information on some meta-paths, the model is able to make recommendations through the associated information on other paths.
Misleading information extraction is also a challenge that needs to be addressed in traditional recommendation models, as they usually model users and items in isolation under each meta-path, resulting in potentially misleading information extraction. The recommendation model based on HIN adopts a unified embedding approach, which describes users and items under different meta-paths through common feature characteristics. This approach reduces misleading information extraction and captures user and item characteristics more comprehensively, thus providing more accurate recommendation results.
Also Read: C3 AI and Shell Expand Collaboration for Asset Monitoring and Predictive Maintenance
When exploring heterogeneous information networks, current traditional recommendation models usually only consider the structural features of the information network, ignoring the potentially useful information in it. The recommendation model based on HIN uniformly embeds users, items and meta-paths into the relevant potential space by learning node embedding vectors. In this way, the model can better quantify the user’s preference for meta-paths, thus improving the effectiveness of personalized recommendations and avoiding the irreversible loss of useful information. WiMi’s HIN-based recommendation model can effectively solve the problems of the current Internet recommendation models and improve the accuracy, degree of personalization, and user experience of recommendations. The model can make full use of the relationships and features in the heterogeneous information network to provide users with more accurate and valuable recommendation results.
The recommendation model based on HIN implementation consists of the following key steps:
Data processing: First, the data in the heterogeneous information network needs to be pre-processed. This includes encoding the representations of users, items, and relationships, e.g., converting them into numerical or vector form for use in the model. Also, a meta-path graph needs to be constructed for describing the relationships between nodes.
Meta-path selection: In HIN, meta-paths are paths describing the relationships between nodes. According to the specific recommendation task and data characteristics, a suitable meta-path needs to be selected. The selection of meta-path should be based on domain knowledge and experience, aiming to capture the relevance between users and items.
Node Embedding Learning: Next, the embedding vectors of the nodes need to be learned to represent the features of users and items under different meta-paths. Embedding learning methods can include deep learning-based methods as well as matrix decomposition-based methods such as matrix decomposition models.
Relationship modeling and feature fusion: in this step, the model uses the learned node embedding vectors to model the relationships between nodes. By considering the interrelationships between meta-paths, feature information under different meta-paths can be fused. Commonly used approaches include using an attention mechanism to model the weights of different meta-paths to better capture the correlations between nodes.
Personalized recommendation: finally, the learned node embedding vectors and relationship modeling results are used for personalized recommendations. By measuring user preferences for different meta-paths, more accurate and personalized recommendation results can be provided. Commonly used recommendation algorithms include content-based recommendation and collaborative filtering algorithms.
SOURCE: PRNewswire
