Project of Team Aelous in Google ML Winter Camp 2020, Shanghai Site

13 - 17 January, 2020

Project Name: Webpage Wizard

Team Members: Muge Chen, Yujie Lu & Yanqiong Chen

1. Project Introduction

In the world of machine learning, people train models based on dataset that are fully labeled. However, most of the application domain suffers from not having sufficient labeled data whereas unlabeled data is available cheaply. And our project also faced the same problem. Specifically, we try to do webpage classification on a dataset that only a small amount of data is labeled.

Here is the basic information of our dataset

• Dataset name: webpage-classification
• Data size: 10K pairs of {web page screenshot image, web page html source}.
• 800 samples are manually labeled {is_entity, category}
• 2590 samples are manually labeled {is_entity}

Our model aims to solve such a challenging issue by semi-supervised learning combined with some rule based methods.

2. Implementation Details

First we did some pre-processing on the dataset.

• We crop and resize the render images to a fixed size of (224,224,3).
• We use a html parser to extract useful information from the html data. And use word2vec to process the extracted information.
• We convert the given labels to probability of different categories.
• We also write a script to label some of the unlabeled data that we are 95% confident. For example, the one with url that starts with wikipedia or contains imdb/title, etc.

Second we have built 3 independent models to classify the web page.

2.1 url model

In this sub model, we use unsupervised cluster algorithm to do classification. However, the category can not be mapped to the clusters we have found, since the dataset is really biased.

2.2 html model

First we use html parser to parse the html into plain text, and next we use doc2vec to transform the text data into a vector of 300 dimension. Then use a deep learning model to predict the category.

It finally reaches an accuracy of 80% on the dataset. However, the model sometimes can not converge, we can see that there are still some problems in the model or the feature engineering.

2.3 image model

We use deep CNN to train the model, and it reaches an accuracy of 81% on the dataset and 95% on the validation set.

2.4 Combined model

In addition, we have put the 2 model together. That is, we use both the html and the render to train the model, and it turns out to perform well on the dataset.

3. Demo and Posters

There are many possible applications of webpage classification, for example

• We can build an extension for Chrome which can beautify the UI with different strategies based on the category of the webpage. Just like switch omega.
• The browser can collect information about which category the user visited most frequently to decide which AD to present.

References

[1] S4L: Self-Supervised Semi-Supervised Learning (ICCV 2019)

[2] Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results (NIPS 2017)

[3] Le, Quoc, and Tomas Mikolov. "Distributed representations of sentences and documents." International conference on machine learning. 2014.