"Recommendation System" learning notes--basic concepts

Actual combat recommendation (development and maintenance of e-commerce site recommendation system) after a few months, feeling reached a bottleneck.

From the practical point of view, for a medium-sized e-commerce site (such as tens of thousands of PV), independent construction of a recommendation system, the completion of the basic recommendations, such as: Viewview (see also saw), Viewwant (see also add car), wantwant (plus car plus car), Collectcollect (collection also collection) such as the recommendation based on collaborative filtering, Similar (guess you like) such as content-based recommendations, from data collection, extraction, calculation, analysis, there is no problem.

However, the understanding of the theory of recommendation, the recommendation strategy and its application scenario, the recommendation algorithm, the recommendation effect evaluation, etc., need to be further improved.

At this stage, read a book, a recommended introductory book: "Recommender system an Introduction", Chinese name is called "recommendation Systems", a few Austrian wrote, Baidu's Lanci translation.

Reading this book, the study of the content of the record, but also write some ideas, to deepen the memory and understanding of the purpose.

Introduction

The book "Recommendation System" focuses on personalized recommendations. Providing a personalized referral system requires the system to know each user's information. The referral system must develop and maintain a user model or user profile to save the user's preferences.

The user model is important for each recommender system, and how to obtain and utilize this information depends on the specific recommendation technology. User preferences can be obtained implicitly by monitoring user behavior, or by asking visitors to display them.

What additional information should I use when generating a personalized referral list? The best-known approach is to consider the behavior, opinions, and hobbies of large groups of other users. Systems are often referred to as group-based or collaborative approaches.

Collaborative filtering recommendations

The basic idea of these systems is that if the user has the same preferences in the past (such as browsing or buying the same book), then they will have similar preferences in the future, such as: User1, User2 similar, User1 bought book1,user2 do not know Book1, You can recommend Book1 to User2.

This technique is also known as collaborative filtering (cf,collaborative Filtering) because the choice of books that may be of interest involves filtering out the best hope from a large set of books, and the user is implicitly collaborating with others.

Frequently asked questions are:

1) How do we find users who have similar preferences to the users we recommend?

2) How to measure similarity?

3) How to deal with new users who have not yet purchased their experience?

4) What if there are very few ratings?

5) In addition to using similar users, what techniques can be used to predict whether a user likes their items?

Pure collaborative filtering does not use or require any knowledge of the item itself, such as the recommendation to sell a book system, and does not need to know the genre, content or author of the book. The obvious advantage of this strategy is that the system does not need to acquire and maintain this data.

Content-based recommendations

The core of content-based recommendation is the ability to get descriptions of items (whether manually generated or automatically extracted) and important catalogs of these features.

Similar to the description of the item, the user record also needs to be automatically extracted or "learned" by analyzing the user's behavior and feedback, or directly asking the user's interests and preferences.

Problems:

1) How does the system automatically acquire and continuously improve user records?

2) How to decide which item to match or at least close to, in line with the user's interest?

3) What technology can automatically extract or learn the description of items, thus reducing manual labeling?

Content-based recommendations have two major advantages:

First of all, there is no need for large-scale users to achieve a moderate degree of recommendation accuracy;

Second, once you get the property of the item, you can immediately recommend new items.

Knowledge-based recommendations

If you focus on other areas of application, such as consumer electronics, it involves a large number of single buyers. This means that we may not be able to rely on a purchase record, which is a prerequisite for collaborative filtering and content-based filtering methods.

Even so, we are able to get more granular and structured content, including the quality features of professionalism.

For example, to buy a digital camera, the average user will buy a new camera every few years, so the recommendation system is not possible to build user records or recommend other people like the camera, which will lead to only recommend the best-selling models.

At this point, the system needs to generate recommendations with additional causal knowledge. In this knowledge-based approach, the referral system typically uses additional information about the current user and the active item (this information is generally provided manually).

In many knowledge-based recommender systems, user requirements must be guided interactively. A truly interactive approach to design should be like a regular conversation, and in a personalized answer, the system can find out what the user really likes.

Overall, the knowledge-based Recommender system solves the following problems:

1) What kind of domain knowledge can be expressed as a knowledge base?

2) What mechanism can be based on the characteristics of users to select and rank items?

3) How can I obtain user information in areas where there is no record of purchase? How to deal with the preference information given directly by the user?

4) which interaction mode can be used for interactive recommender systems?

5) What personalization factors should be considered when designing a dialogue to ensure accurate user preference information?

Evaluation recommendation System

The main driving force in the field of recommendation systems is to improve the quality of recommendations. The question that follows is how do we actually measure the quality of the recommendations that the Recommender system gives? What assessment methods should we choose?

Answers to the following questions can be answered:

1) which research designs are suitable for evaluating recommender systems?

2) How to use historical data to evaluate the recommendation system?

3) What measurement criteria are appropriate for different assessment objectives?

4) What are the limitations of existing assessment techniques? Especially in terms of conversational or commercial value of the referral system.

Summary

This chapter is a list of some of the basic concepts that are relevant to the recommendation system and will gradually deepen the learning of each part in a later chapter.

For the referral system, you can answer the following questions:

1) What is the business value of the referral system?

2) can it help increase sales or convert many visitors to buyers?

3) Are there any differences in the effectiveness of the recommended algorithms? What kind of technology should I use?

"Recommendation System" learning notes--basic concepts