Commodity sequence recommendation method based on deep learning

Abstract

The invention discloses a commodity sequence recommendation method based on deep learning, and the method comprises the following steps: 1, obtaining a public commodity sequence purchased by a user, carrying out the preprocessing of the commodity sequence, and obtaining commodity names and the metadata of commodities; 2, taking the commodity names and metadata as input and transmitting the commodity names and the metadata to a designed sequence recommendation model, and obtaining a pre-training model, wherein the pre-training model is divided into an input layer, a coding layer, a multi-layerperceptron layer and a mapping layer; 3, selecting proper output on an output layer of the pre-training model as representation of a current sequence, performing further calculation to obtain a recommendation commodity list. According to the method, the two-way influence of daily shopping commodities is considered, an encoder-decoder structure is adopted, two-way modeling can be carried out on a commodity sequence, encoding information acquisition is superior to that of a one-way model, and final recommendation content can be optimized.

Description

technical field [0001] The invention relates to the technical field of fault handling, in particular to a method for recommending product sequences based on deep learning. Background technique [0002] With the continuous development of e-commerce platforms, the number of commodities is growing exponentially; the expansion of commodity scale, on the one hand, increases the possibility of meeting the different needs of users, but on the other hand, it also intensifies the need for users to locate effective commodities from a large number of commodities. time cost. There are hundreds of millions of products in e-commerce platforms such as Taobao and Amazon, and users often need to spend a lot of time searching for the products they are interested in. Users are usually only interested in a few items when shopping online, and a large number of irrelevant redundant items seriously affect the normal search for the items they are interested in. Due to time and resource constraint...

AI Technical Summary

Problems solved by technology

[0013] The disadvantage of the existing technology is that the non-sequential recommendation method is mainly unable to completely simulate the dynamic shopping process of the user, and usually only predicts the possible preference of the user or performs top-n recommendation, which is inconsistent with our actual situation

And RNN is not good at modeling long sequences, and its weight distribution is often based on the nearest neighbor nodes

The sequence recommendation method based on convolutional neural network (Jiaxi Tang et al. proposed Caser) cannot efficiently capture long-term dependencies due to the size limitation of convolutional filters. At the same time, due to the limitations of CNN itself: local links, weight sharing, so Not as good as RNN in text processing

The transformer-based sequence recommendation method (SAS4Rec proposed by Kang W et al.) only considers the front-to-back relationship of shopping and ignores the two-way impact of shopping, and does not consider other attributes of the product

To sum up, the existing technology cannot perform two-way modeling on commodity sequences, which limits the ability of implicit representations in historical sequences, and most of them are based on RNN recommendation methods, which cannot establish long-term dependencies and have serious weight loss. The problem is that the recommended products are often based on the recommendation results obtained at the end of the shopping sequence, ignoring the influence of the sequence head

Images