This paper describes our taxonomy classifier for SIGIR eCom Rakuten Data Challenge. We propose a taxonomy classifier based on sequenceto-sequence neural networks, which are widely used in machine translation and automatic document summarization, by treating taxonomy classification as the translation problem from a description of a product to a category path. Experiments show that our method can predict category paths more accurately than baseline classifier.
• Computing methodologies → Information extraction;
Taxonomy is the major classification schemes in organizing concepts. With the rapid growth of the e-commerce market accompanying the development on the Internet, the number of products on e-commerce becomes enormous. In this situation, it is required to develop methods that predict taxonomic categories automatically because it is costly to classify all the products manually.
Rakuten Data Challenge, which is a competition we participated, provides a task to predict correct categories for each given product. As a feature of this task, categories have a hierarchical structure. This hierarchical structure corresponds to a taxonomy, which indicates that items in a category are further classified into a subcategory that contains further lower detail information. Each product has a path in the taxonomy like “Clothing, Shoes & Accessories → Shoes → Men → Boots”.
As an approach to solving this task, the most straightforward approach is to train a multi-class classifier (e.g., Random Forest) that predicts a category path as a class of a given product. However, Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s).
SIGIR 2018 eCom Data Challenge, July 2018, Ann Arbor, Michigan, USA © 2018 Copyright held by the owner/author(s).
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. https://doi.org/10.1145/nnnnnnn.nnnnnnn as mentioned earlier, the number of category paths is 3,695, which is fairly large to be considered as a set of classes for ordinal machine learning classifier. Moreover, this approach independently treats these category paths although a category path shares a part of another category path of a similar product. It is expected that this fact causes more data sparseness issue and degrades the performance because the classifier has no way to find common patterns that are shared in two diferent category paths.
In this paper, we propose a taxonomy classifier based on EncoderDecoder neural networks. The key idea is to regard the category path as a series of category names in each hierarchical level. From this perspective, the taxonomy classification task can be converted into a sequence-to-sequence problem, which has a text (i.e., a sequence of words) of the product name as the input and a sequence of category names as the output. In recent years, remarkable performance has been demonstrated in the field of machine translation and automatic summarization by using the model called neural network Encoder-Decoder architecture. We apply the EncoderDecoder model to the taxonomy classification task and evaluate the performance. Experiments show that our approach can successfully predict category paths more precisely than the baseline approach that treats the task as a multi-class classification problem and applies Random Forest. 2
We have 800,000 records for training data and 200,000 records for test data. Each record has a description of a product and a category path. The number of labels in the training data is 3,695, and each label is assigned to 868 items on average. The category (id=4015) is most frequently assigned to products, which is assigned to 268,295 items. Figure 1 shows the histogram of the number of words in each description in the training data. The average number of words was 10.92, and the standard deviation was 5.21. The maximum number of words was 58, and the minimum value was 1.
20 30 40 number of words in product descriptions in the dataset 50 60 (2) (3) (4) (5)
We used 20 % of the training dataset as the validation set to evaluate models. As preprocessing, we lowercase a product name in training/validation/test sets with SpaCy 1. We use both the original corpus and the lowercase corpus and compare classifier performances.
For the weights of dense word representation layer, we use
GloVe [
Encoder-Decoder Neural Network is a type of neural network that
is actively studied in recent years [
Figure 2 shows our Encoder-Decoder neural network with
attention mechanism [
i and attention is defined as following:
aˆt i ,
at i = Íj aˆt j
aˆt i = att (hi , h¯t ),
att (hi , h¯t ) = hi T Wah¯t ,
matchinд rate = |VDat aset ∩ VGloV e | ,
|VDat aset |
(1)
where att (ht , h¯i ) is an attention function. The attention function
of our works is based on Luong et al. [
After the encoder takes input, the decoder predicts outputs using encoder state. As a feature of the Encoder-Decoder neural network, the input sequence length and the output sequence length do not have to match. It can predict various length category path with various length of a description of a product. 4
This section presents evaluations of our taxonomy classifier and the baseline classifier in the validation set. At the time of training, we use up to 50,000 words as the features both in baseline and the proposed model. In the experiment, we examine parameters of our taxonomy classifier (in Table 3) and show best parameters in each pair of encoder and decoder in Table 4. 4.1
We use Random Forest [
We evaluate the performance of our proposed models and the baseline on the validation set with the oficial script ( eval.py). We show the best parameters for each model in Table 4, and the results in Table 5. Bidirectional LSTM with GloVe achieves the best F1 score. Our model achieved the best performance when it uses Bidirectional LSTM as an encoder/decoder, lowercase dataset and use GloVe embeddings to initialize the weights of the embedding layer for the input sequence. Interestingly, it shows bad scores when we use GRU for encoder and decoder. We will further investigate the reason for this. 5
In this paper, we propose an encoder-decoder neural network for taxonomy classification where there are various sizes of category paths. It is computationally expensive to solve this problem as a multi-label classification because there are over 3,695 categories in the dataset, To avoid this dificulty, we regarded taxonomy classiifcation as the translation from the description of products to the Word embedding dim for source (product) Pre-trained word embedding Word embedding dim for target (category) Encoder / Decoder units Number of Encoder / Decoder Encoder / Decoder dim Global attention [LSTM, BiLSTM, GRU, BiGRU] GRU LSTM 300 500 300 500 encoder and decoder.
RNN dim
Embedding type
GRU
This work was supported by JSPS KAKENHI Grant Number 16H02904. Also, we would like to show our gratitude to Kento Nozawa and Taro Tezuka for comments that greatly improved the manuscript. 0.5367 0.6031