MRNet-Product2Vec: A Multi-task Recurrent Neural Network for Product Embeddings

Abstract

Abstract. E-commerce websites such as Amazon, Alibaba, Flipkart, and Walmart sell billions of products. Machine learning (ML) algorithms involving products are often used to improve the customer experience and increase revenue, e.g., product similarity, recommendation, and price estimation. The products are required to be represented as features before training an ML algorithm. In this paper, we propose an approach called MRNet-Product2Vec for creating generic embeddings of products within an e-commerce ecosystem. We learn a dense and low-dimensional embedding where a diverse set of signals related to a product are explicitly injected into its representation. We train a Discriminative Multi-task Bidirectional Recurrent Neural Network (RNN), where the input is a product title fed through a Bidirectional RNN and at the output, product labels corresponding to fifteen different tasks are predicted. The task set includes several intrinsic characteristics about a product such as price, weight, size, color, popularity, and material. We evaluate the proposed embedding quantitatively and qualitatively. We demonstrate that they are almost as good as sparse and extremely high-dimensional TF-IDF representation in spite of having less than 3% of the TF-IDF dimension. We also use a multimodal autoencoder for comparing products from different language-regions and show preliminary yet promising qualitative results.

P-Companion

“…In terms of type level, Comparing to JOIE with the item-type view, P-Companion improves by 9.9% on the Electronics dataset and by 4.5% on the Grocery dataset. We observe similar phenomena on the larger "All-Group" dataset with a relative 3.8% increase on itemlevel Hit score on average and increase on type-level against JOIE 3 . We believe this is due to the following reasons, (i) P-Companion infers complementary products by targeting the complementary type first rather than only modeling product relationships in Sceptre and PMSC.…”

“…Current methods [17,23] often fail on such low-resource products, which widely exist in e-commerce. While most of the existing methods in recommender systems [21] focus on modeling user-item relationships by frequent pattern mining [8], matrix factorization [15], collaborative filtering [14,20], or other neural network based recommenders [2], only a few [3,8,10,16] target at explicitly modeling relationship between items. Among them, complementary relationship modeling has been scarcely investigated compared to the efforts made for modeling substitutes with similarity-based approaches.…”

“…where is the feature vector for item 's catalog and activation function (•) = tanh(•). (1) , (2) ∈ R × , (3) ∈ R × are weight matrices with the corresponding biased terms (1) , (2) , (3) . Figure 4: GNN-based Product2vec module architecture, which learns effective product embeddings given its textual features and aggregation from similar products.…”

“…Although historical co-purchase data is good as distant supervision labels, it is far from complete and fails to include all possible truths on complements. Therefore, we leverage MTurk to evaluate the 3 Due to the scalability issue, the results of PMSC is not available. performance of P-Companion with an open-world assumption.…”

Large-scale Collaborative Filtering with Product Embeddings

“…Reusing a shared set of highly informative pre-trained item embeddings significantly reduces computational costs, and allows us to focus our efforts on experimentally measuring the effectiveness of downstream collaborative filtering models. While there has been a number of item embedding algorithms proposed in the literature [Grbovic and Cheng, 2018, Grbovic et al, 2015, Biswas et al, 2017, Wu et al, 2017, Vasile et al, 2016, to the best of our knowledge, methods of leveraging such pre-trained item embeddings for user preference modeling have not been extensively studied.…”