you may also like
Today in doing the project, I found a particularly interesting module, called guess you like it,
give us a presentation mall recommendations and you may also like is how to achieve, first you have to know, you recommend a product, certainly not just give you a push, through algorithms that are then displayed to your recommendation area
Why do I need recommendation system?
Because of information overload, people get information from the active into a passive access to information, open the phone every morning ......., Often effective access to data is extremely important, all targeted demand will arise naturally, the recommendation system applications students.
The first recommendation system because of the Amazon is to sell goods produced, the large number of goods on Amazon's online sales, but few in the home display of goods.
Amazon's product display page, often see: Browse 00 users to simultaneously browse *** commodity, this is a very simple recommendation.
Cold start problems:
Recommended system must have a huge data as a support, but when we restart the product, not a lot of user data, this time will encounter cold start problems,
Cold start to solve the problem:
Since there is no pre-user data, then we can use the data to recommend the product itself, such as the price of the same commodity, the same brand of goods,
Content-based recommendation:
1, to tag the goods: food, sports, clothes, etc. The more detailed the better, the more detailed the more accurate.
However, according to these commodities recommendation will go wrong tonight Xiaoming search Ferrari models in our products, our system tomorrow morning by Ferrari this keyword to Bob recommended a five million Ferrari.
Of course, the use of product recommendations but no user data is preliminary to use, and at a later stage, more and more users, we want to start using the user behavior data.
User behavior data recorded in the end what?
Xiao Ming opened the Amazon browser, click a few hot suit, then search bar to search the nike shoes, after reading a few pairs of shoes 10, he bought a pair of shoes Anta.
This is the user behavior data. The more of this data, we can do greater things.
Weighting algorithm:
User-time user preference matrix: This is where the user happens to view, purchase, share the behavior of commodities. These actions are varied, it is necessary to calculate the weighting algorithm preference degree of the user for an article, then a user-item user preference matrix.
After the user behavior will generate noise and user errors. We got the data, the data should go wash, the most important is that the noise reduction and normalization.
Noise Reduction: For example, because of a network outage, users in a short time had a large number of clicks of the operation. Through a number of strategies and data mining algorithms to remove noise in the data.
Normalization: The purpose is to clean up the data by weighting the different behavior, to form a reasonable user preference matrix. The simplest algorithm is normalized by dividing the maximum value of the various types of data to such data in order to ensure that all data ranges are within [0,1].
Dimensionality reduction algorithm --SVD singular value decomposition
By user behavior more and more data, more and more sparse data at this time, fewer and fewer valid data at this time, SVD algorithm was born, as far as I know, Taobao is with this recommendation
Made torture more than pave the way, it is to make it clear synergy algorithm. Algorithm commonly known as the co-SF, now widely used in today's recommendation system, through cooperative algorithm,
Why is called collaborative filtering? Because these two matrices are calculated by the other party.
Target collaborative filtering algorithm (collaborative filtering) is based on user comments on the articles of history, it is recommended not to buy items to target users (active user). Based collaborative filtering algorithms can be divided into objects, and based on user-based matrix decomposition, and the articles herein Based on collaborative filtering algorithm based on matrix decomposition.
Overview collaborative filtering algorithm
Input data: Typical problems collaborative filtering input data of m users list, n a list item, each list item while a user has been evaluated.
Recommended process:
1. Calculation of Similarity (Item similarity between users or between): usually pearson correlation coefficient or cosine similarity.
2. Calculate prediction (not target users buy attractive items to target users): The target user for prediction score did not purchase goods.
3. Calculate recommendation: Recommended biggest attraction of the first K objects to the target user.
Based on user collaborative filtering algorithm
Collaborative filtering algorithm based on the user's aimed at finding similar users, and then recommend similar items between users.
1. similarity: calculating the similarity between users. When it comes to above have been evaluated for each user has a list of items, then this list is a property vector of the user, the user's similarity is the similarity between the vector.
2. prediction：假设用户A和B、C是相似用户。假设Item1, Item2, Item3三个物品是B、C购买过但A未购买过的物品。那么我们就可以向A推荐 这些物品。如何计算这三个物品对用户A的吸引力呢？以B、C和A的相似度为权重，计算B、C对物品的评分均值即可。
基于用户的协同过滤算法实际上面临很大的问题，例如稀疏性问题，毕竟一个用户购买的物品是非常少的。
关于代码特别复杂,放到这里也不好理解 .大家可以去晚上搜一下 :: 协同过滤的代码,再细细理解.