Participation in the award: This article has participated in the "Newcomer Creation Ceremony" activity, together with the start of the road of gold digging creationCopy the code
Thesis Title:UGRec: Modeling Directed and Undirected Relations for Recommendation
Presentation: SIGIR2021
A, Motivation

Knowledge graph has attracted a lot of research because it can encode the relationship between user and commodity, but little attention is paid to the co-occurrence information between commodity and commodity, which contains abundant similar information between commodity and commodity. To further illustrate the importance of co-occurrence information, the author gives the following example:

Item 2 is an action/adventure game made by Capcom, and Item 3 is an RPG made by Square Enix. These two games belong to different manufacturers, different genres. There is no clear attribution information between the two, but consumer trends suggest that both are often bought together by video game enthusiasts. This shows that there are some common features between the two games that can arouse consumer interest while buying them. But this relationship is difficult to observe in the knowledge graph (limitations of using only the knowledge graph). Item 4 and Item 2 belong to the same type of game, but the lack of Item 4 in the data set the category information, so that we can’t through the category information connects two Item by category, but between them they have a common comment on information, this can help to study the relationship between them and thereby help alleviate the problem of the missing category information.

Second, the Model

The model is mainly composed of two parts. The Central Entiry Space is modeled by existing models. On the lower left is a plate dealing with indirect co-occurrence information, the core of which is to map the indirect relationship between item and item to different hyperplanes. At the top right is a section that deals with items and users that are directly related, as well as an attention mechanism for dealing with finer grained relationships.

Direct relation space: HP, TP and RP represent different projection vectors of <head, tail and Relation > respectively. Ikxk represents a square matrix of K dimension and the four embedding Vectors can be obtained through model learning. Mrhh and Mrtt are mapping matrices for HRD and TRD (not randomly initialized matrices). HRD and TRD represent the projection embedding of head and tail entities in Relation R. W is the trainable weight matrix and B is the deviation vector. The loss function uses the standard Hinge loss. S and S(underlined) represent positive and negative samples respectively. M >0 indicates the safe edge size.

Indirect contact space: a matrix based learning approach. Hc and TC are projected onto the hyperplane of R, and the attention mechanism is used to capture fine-grained indirect relationships between entities, and then the distance between them in the hyperplane is minimized. Hinge loss functions are used as loss functions for directly related data. Lambdar and LAMBdac are hyperparameters used to control the influence of direct and indirect connections on modeling.

Data & Experiments

Data sets: Three data sets (public data sets)

Four, the Performance of

It can be seen from the final performance that the proposed model achieves SOTA effect in all three data sets

Five, the Ablation Study

Ablation experiments were compared from top to bottom: only user-item interaction information was included; Contains only data that is directly related; Contains only data that is not directly related; The results with direct and indirect data but without attention mechanism verify the effectiveness of adding direct and indirect data into the model.

Six, the Conclusion

The superior performance over two recent KG-based recommendation models validates the effective design of UGRec on separately modeling the directed and undirected relations.

The final performance demonstrates the effectiveness of UGRec for direct and indirect linkage modeling.