面向金融产品推荐的协同过滤算法研究

Research on Collaborative Filtering Algorithms for Financial Products Recommendations

杨博

11749123

硕士

信息与通信工程

骆宗伟

2019-05-29

2019-07-15

哈尔滨工业大学

深圳

随着经济社会的不断发展和金融市场的不断活跃，对于金融产品的推荐成 为一个不可忽视的问题，而股票作为一种非常重要的金融产品，尤其值得关注。 由于我国的股票市场起步较晚，依然存在着很多的不足之处，股民普遍存 在一些关于股票推荐和选股方面的问题。具体表现在一是当前股民投资的理论 基础不强，缺乏金融投资的基础知识，其投资行为往往只是一个模糊的操作， 不具有很强的可解释性，甚至投资者自身也很难说清楚自己的投资逻辑，很大 一部分投资者投资的最大的理由往往是“感觉”这么一个说不清道不明的东西。 二是现存的一些成熟的股票推荐模型其主要是以股评的方式作为支撑，具有相 当笼统和模糊的特征，有一定的局限性。三是往往在构建推荐模型时候,特征之 间存在相互关联，导致模型的复杂度较高，存在计算和存储上的困难，如何在 保留主要信息和不损失模型精确度的情况下同时降低模型的复杂度是一个值得 研究的问题。综上，所以当前急需一个能够精准刻画用户肖像的模型来为广大 股民服务。这个模型应当能对股民的投资行为作出合理且清晰的解释，能让股 民以直观通俗方式的理解，另外在模型的便捷，方便性上也要存在一定的优势， 因为对于一般普通的中小股民来说，复杂的模型过于的艰难晦涩，不方便使用， 那么这种推荐方法在可行性上就大大的打了折扣。 针对存在的问题,考虑到协同过滤算法在商品的推荐应用中具有巨大的优 势和比较成熟的应用模式。受此启发，本文考虑将协同过滤算法融合到金融产 品的推荐中，而股票作为一种常见的金融产品，对于协同过滤算法的应用有着 得天独厚的优势。将股票当作商品，股票的多种指标当作物品的多种特征，将 客户选择股票比作客户选择物品，给客户推荐股票当做给客户推荐物品，股票 的指标值拟化为顾客对物品的打分。最后对于应用多指标因子建立模型时导致 模型复杂度过高，并且影响模型的计算和存储效率的问题，本文通过主成分分 析（Principal Component Analysis, PCA）降维的方法在保留绝大部分信息，不 影响模型精度的情况下来对模型进行有效的降维，然后从余弦相似度和平均绝 对误差两个方面得出一个最佳的降维维度，使得生成的推荐模型更加的简约， 轻量，精准和便捷。实现了一个科学的股票类金融产品推荐模型。 这种模型在理论上有一定高度，想法上有一定的创新性，在实际上也具有 较强的实践性和可操作性。

With the continuous development of economy and society and the active financial market, the recommendation of financial products has become a problem that can not be ignored. As a very important financial product, stock deserve special attention. Due to the late start of Chinese stock market, there are still many shortcomings of the stock market. There are widespread questions about stock recommendation and stock picking among investors. The specific performance is that the investors' current investment theoretical is not strong, and the basic knowledge of financial investment is lacking. The investment behavior is often only a vague operation, and its behavior is not very interpretable. Even investors themselves are hard to explain the logic of their investment, the biggest reason for a large part of investors to invest is often due to their "feeling" ,something that is unclear. Second, some existing stock recommendation models are mainly supported by the way of stock evaluation and comments. They have quite general and vague characteristics and have certain limitations. Thirdly, when constructing the recommendation model, the features are related to each other, resulting in high complexity of the model, difficulties in calculation and storage, how to preserve the main information and without losing the accuracy of the model and reduce the model's complexity is a problem worth studying. In summary, there is an urgent need for a model that accurately portrays user portraits to serve the broad masses of investors. This model should be able to make a reasonable and clear explanation of the investor's investment behavior, allowing the investors to understand in an intuitive and common way, and also have certain advantages in the convenience of the model, because for ordinary investors.The complicated model is too difficult and inconvenient to use, and this recommendation method is greatly discounted in terms of feasibility. For the existing problems, it is considered that the collaborative filtering algorithm has great advantages and relatively mature application modes in the recommended application of commodities. Inspired by this, this paper considers the integration of collaborative filtering algorithms into the recommendation of financial products.As a common financial product, stock has a unique advantage for the application of collaborative filtering algorithms. Just look the stocks like the commodities, and various indicators of stocks as a variety of characteristics of the item, the invsetor selects the stock as the customer select the commodities, and recommends the stock to the customer as the recommended commodities to the customer. The index value of the stock is calculated as the customer's score on the commodities. Finally, the model's high complexity caused by using multiple indicator factors to build the model, which also affects the calculation and storage efficiency of the model. In this paper, across to use the Principal Component Analysis (PCA) dimensionality reduction method preserves most of the stock information, and effectively reduces the dimension of the model without affecting the accuracy of the model, and then depend on the cosine similarity and average absolute error lead to an optimal dimension reduction dimension. The generated recommendation model which is more simple, lightweight, precise and convenient. Ascientific stock financial product recommendation model was implemented. This model has a certain height in theory, and it has certain innovation in its ideas. In fact, it also has strong practicality and operability.

股票 推荐系统 PCA 相似度 协同过滤

stock recommendation system PCA similarity collaborative filtering

中文

联合培养

南方科技大学

杨博. 面向金融产品推荐的协同过滤算法研究[D]. 深圳. 哈尔滨工业大学,2019.