Discussion 1

YouTube Recommendor system

As the Internet continue to evolve, information and data is unlimited and people become overwhelmed and confused about what they are looking for. With the help of recommender systems, a lot of useful information can be filtered out to help users discover and select the information that is of interest to them. There are thousands of videos available on YouTube and it may be difficult for users to find videos that best match their interests. YouTube recommendations are responsible for helping more than a billion users discover personalized content from an ever-growing corpus of videos. In this paper we will focus on the immense impact deep learning has recently had on the YouTube video recommendations system.

Recommending YouTube videos is extremely challenging from three major perspectives:

Scale, Freshness, Noise

YouTube Recommender system is built on Google Brain which was recently open sourced as TensorFlow. TensorFlow provides a flexible framework for experimenting with various deep neural network architectures using largescale distributed training. Neural networks are used for recommending news, citations and review ratings. Collaborative filtering is formulated as a deep neural network in and autoencoders. It uses deep learning for cross domain user modeling. And in a content-based setting,used deep neural networks for music recommendation. As a complex deep learning model youtube recommender system uses both content based and collaborative filtering techiques.

System Overview

The overall structure of our recommendation system is illustrated Below.

The system is comprised of two neural networks: one for candidate generation and one for ranking. The candidate generation network takes events from the user’s YouTube activity history as input and retrieves a small subset (hundreds) of videos from a large corpus. These candidates are intended to be generally relevant to the user with high precision. The similarity between users is expressed in terms of coarse features such as IDs of video watches, search query tokens and demographics.

The ranking network accomplishes its task by assigning a score to each video according to a desired objective function using a rich set of features describing the video and user. The highest scoring videos are presented to the user, ranked by their score. The two-stage approach to recommendation allows to make recommendations from a very large corpus (millions) of videos while still being certain that the small number of videos appearing on the device are personalized and engaging for the user. Furthermore, this design enables blending candidates generated by other sources, such as those described in an earlier work.

Avoiding Attacks

In order to aid against the exploitation/exploration problem, youtube feed the age of the training example as a feature. This helps overcome the implicit bias in models which tends to recommend stale content, as a result of the average watch likelihood during training time. At serving time, they simply set the age of the example to be zero. In addition videos too similar to each other are removed, and the number of videos coming from the same channel is limited. For multi-armed bandits the upper confidence bound could be applied. The idea of upper confidence bound is using the upper bound of the estimated success rate (e.g., order-rate, click-rate, watch-rate). When a new item arrives without any information, the confidence interval is wide, and so the upper bound is high. Therefore, the new item would have a high chance of being recommended. As the item gets more impressions, the estimation would be more accurate, and the upper bound would be closer to its actual value.