Substitutable/complementary product recommendation
When a user has browsed a t-shirt, it is reasonable to retrieve similar t-shirts, i.e., substitutes; whereas if the user has already purchased one, it would be better to retrieve trousers, hats or shoes, as complements of t-shirts. In [Wang 18], for each product, its embedding representation is learnt in a general semantic space. Thereafter, embedding vectors are projected into two separate spaces via a novel mapping function. In the end, each embedding is incorporated with path-constraints to further boost discriminative ability of the model.
Given a network of items, the objective in [Rakesh 19] is to learn their content features such that they explain the relationship between items (e.g., Xbox) in terms of substitutes (e.g., PS4) and supplements (game controllers, surround system, and travel case). A generative deep learning model links two variational autoencoders using a connector neural network to create Linked Variational Autoencoder (LVA) in this regard which in turn is extended by incorporating collaborative filtering (CF) to create CLVA that captures implicit relationship between users and items.
Attribute-aware collaborative filtering (A2CF) [Chen 20], instead of directly modeling user-item interactions, extracts explicit and polarized item attributes from user reviews with sentiment analysis, whereafter the representations of attributes, users, and items are simultaneously learned. Then, by treating attributes as the bridge between users and items, user-item preferences (i.e., personalization) and item-item relationships (i.e., substitution) are modemed for recommendation. In addition, A2CF is capable of generating intuitive interpretations by analyzing which attributes a user currently cares the most and comparing the recommended substitutes with her/his currently browsed items at an attribute level.
Unlike simple relationships such as similarity, complementariness is asymmetric and non-transitive. Standard usage of representation learning emphasizes on only one set of embedding, which is problematic for modeling such properties of complementariness. Contextual knowledge is encoded into product representation by multi-task learning, to alleviate sparsity in [Xu 20]. By explicitly modeling with user bias terms, noise of customer-specific preferences is separated from complementariness. Furthermore, dual embedding framework is adapted to capture the intrinsic properties of complementariness and provide geometric interpretation motivated by classic separating hyperplane theory. Finally, a Bayesian network structure unifies all the components.
领英推荐
Product relations and user preferences are modeled with a graph attention network and a sequential behavior transformer, respectively in [Yan 22]. The two networks are cast together through personalized re-ranking and contrastive learning, in which the user and product embedding are learned jointly in an end-to-end fashion. The system recognizes different customer interests by learning from their purchase history and the correlations among customers and products.