Recommendation system knowledge (1): Introduction to recommendation system

Introduction

1. What
   a. User: Recommendation system is a tool to help users quickly find useful information
   b. Company: Recommendation system is a tool to increase the probability of a company’s products contacting users, buying, etc.
2. Why
   users : Filter information when the user needs are not very clear. Compared with the search system, the recommendation system makes more use of the user's various historical information to guess the content they may like.
   Company : Solving the product can maximize Attract users, retain users, increase user stickiness, and increase user conversion rate, so as to achieve the goal of continuous growth of the company's business goals. In essence , it is a means to maximize the benefits between users, products and companies.
3. Who
   can see from 1 and 2 above that users and companies are the main objects that need a recommendation system. Then you can start on the basis of 1 and 2 to think about what kind of people need a recommendation system, and what kind of company needs a recommendation system.

In fact, the recommendation system is a system that assists the effective exchange of information between "customers" and "merchants" on various interactive platforms. On the one hand, it helps customers find the products they want and need, and on the other hand, it also helps Merchants find suitable sales channels and audiences, and their core role is to quickly provide effective information to both parties in need and speed up the conclusion of transactions.

Common evaluation indicators

1. Actual feedback-user satisfaction
   Users are very important participants in the recommendation system, and their satisfaction directly determines the quality of the recommendation system. However, the indicator of user satisfaction cannot be
   calculated offline , only through user surveys or online experiments Obtained. Online experiments here are generally obtained through user's online behavior statistics
   1. Customer user feedback: customer's evaluation of recommended content, click-through rate of recommended content, customer stay time and conversion rate, etc.
   2. Feedback from business users: business sales, business profitability, whether the business is willing to pay for it, etc.
2. Quantitative index-prediction accuracy The
   prediction accuracy is used to measure the actual behavior of the user and the accuracy of the prediction result of the recommendation system. This indicator is the most important offline evaluation indicator because it can be
   calculated offline. The following is the most accurate prediction Two commonly used indicators.
   1. Score prediction:
      Predicting the user's scoring behavior on an item becomes a scoring prediction.The scoring prediction model models the user's historical item score records to obtain the user
      's interest model, and then uses the model to predict the user's unseen product score. The prediction accuracy of rating prediction is generally calculated by root mean square error (RMSE) and
      mean absolute error (MAE). For a user and item in the test set, let the user's actual rating of the item, but
      the rating predicted by the recommendation model , Then RMSE can be defined as:
      MAE is defined as:
      RMSE has a square term, which makes users who have a large difference between the user’s real score and the recommendation system’s predicted score increase the penalty, that is, the evaluation index is
      more demanding on the system
   2. TopN recommendation: When the
      recommendation system recommends items to users, it will often give users a list of recommended items. The recommendation in this scenario becomes TopN recommendation. The
      most commonly used prediction accuracy index for this recommendation method is generally precision (precision ) And recall rate (recall), let the recommendation list obtained through the recommendation model,
      be the user's behavior list in the actual scene (test set).
      a. Precision: The number of correctly classified positive samples accounted for the classifier The proportion of the number of samples judged as positive samples (here equivalent to
      the positive samples judged by the model)
      b. Recall: the proportion of the number of correctly classified positive samples to the number of true positive samples (here equivalent to true the positive sample set)
      sometimes for a more comprehensive assessment TopN recommendation, usually choose a different list of recommended calculating the length of multiple sets of precision and recall were then plotted
      precision and recall rate curve curves should be noted that not here PR curve, if you are interested, you can learn about the knowledge of PR curve.

 

Common technology