Life-Long Machine Learning for Recommender Systems

我们人类非常善于做决定。在建立用于决策的自动化方法时，包括数学家和人工智能工程师在内的计算机科学家从人类那里获得灵感。特别是，我们正在努力建立达到人类水平的计算机系统。如果这种方法成功，它们可以帮助人类做出更好的决定。这在我们面临大量选择的情况下或在我们不知道每个选择的可能结果的情况下可能特别有用。例子包括针对特定疾病的治疗选择的决定，或者更普通的决定，比如从一大堆书中选择下一本书阅读，或者选择一组股票添加到我们的金融投资组合中。 人工智能研究中的一种特殊方法是将这样的问题框定为推荐系统。推荐器系统的核心是在用户行为中发现模式，以改善个性化体验，并了解他们所处的环境或背景。最简单的推荐系统根据用户过去的行为向他们推荐感兴趣的项目(例如，下一本要读的书或到目前为止的人)。 虽然推荐系统是当前研究的一个重要领域，也得到了广泛的应用(想想谷歌的搜索建议或亚马逊的“购买了这件商品的顾客也购买了”)，但它们的能力有限。对人类来说，至关重要的是我们有能力根据我们的偏好如何随着时间的推移和环境的变化来调整我们的决定。例如，如果开始下雨，我当天的计划将如何改变。同样，我们人类可以在更长的时间内进行推理。例如，我如何根据我的价值观和未来的就业机会选择大学专业。推荐系统和更广泛的人工智能还没有提供同样的可能性。特别是，推荐系统中使用的数学模型不具有终身学习的能力，即连续地对不断变化的用户偏好和上下文进行建模，并相应地调整其推荐。 这项研究项目将开发和验证新的数学工具，使用机器学习方法在推荐系统中进行终身学习。具体地说，我的研究将在快速增长的深度学习和强化学习(RL)领域使用和开发新技术。深度学习可以为用户随时间推移的行为提供更准确的数学模型，还可以结合不同的信息来源。RL提供了将模型洞察转化为好的(序列)决策的方法。 这项研究将有助于加拿大在人工智能和燃料下游应用方面的领先地位，例如可以帮助加拿大人日常生活的自动健康助手。

We, humans, are very good at decision making. When building automated methods for decision making, computer scientists including mathematicians and engineers in artificial intelligence take inspiration from humans. In particular, we are trying to build computer systems that reach human-level performance. If such methods were successful they could assist humans in making even better decisions. This could be particularly useful in situations where we are faced with a a large number of choices or situations where we do not know about the possible outcomes of each choice. Examples includes decisions about treatment options for a particular medical condition or, more mundane ones, like choosing the next book to read from a large catalogue of books or a group of stocks to add to our financial portfolio. One particular approach in artificial intelligence research is to frame such problems as a recommender system. At their core, recommender systems find patterns in user behaviour to improve personalized experiences and understand the environment, or the context, that they are acting in. The simplest recommender system suggests items of interests to its users based on their past behaviour (for example the next book to read or person to date). While recommender systems are an important field of current research and are widely deployed (think of Google's search suggestions or Amazon's "Customers Who Bought This Item Also Bought") their capabilities are limited. Of crucial importance to humans is our ability to adapt our decisions based on how our preferences evolve through time and based on changes in the environment. For example, how will my plans for the day change if it starts raining. Similarly, we, humans can reason over longer periods of time. For example, how do I pick a university major based on my values and future employment opportunities. Recommender systems and artificial intelligence more broadly do not yet offer the same possibilities. In particular, the mathematical models used in recommender systems do not have the ability for life-long learning, that is to continuously model changing user preferences and context and, accordingly, to adapt its recommendations. This research project will develop and validate novel mathematical tools using machine learning methods for life-long learning in recommender systems. Specifically, my research will use and develop novel techniques in the fast-growing fields of deep learning and reinforcement learning (RL). Deep learning can provide more accurate mathematical models of user behaviour over time as well as incorporate different sources of information. RL provides methods for turning model insights into good (sequences of) decisions. This research will contribute to Canada's position as a leader in artificial intelligence and fuel downstream applications, such as automated health assistants, that can help Canadians in their daily lives.