ChatGPT: Breaking the "cage" of recommendation algorithms

"Guess what you like" why didn't we guess our preferences? Because existing recommendation algorithms have no "perception", they can only use the logic of analogy to analyze what we may want. The LLM (Large Language Model) led by ChatGPT can understand human language, and it can better "see" the true intention behind a sentence, that is, it has stronger perception ability, which can just make up for the shortcomings of existing recommendation algorithms.

Perhaps, the current recommendation algorithm is not so intelligent and "understands me":

• The system always likes to recommend popular products, but "I" is a maverick who likes niche products;
• The system always likes to recommend based on historical purchase lists, but "I" do not need to buy duplicate or similar products;
• The system always needs to enter clear commodity requirements, but "I" often don't know what I want;
• And the "guess you like" who claims to be intelligent, "I" is not so like.

For example, the same user may like the same type of goods (users who love to watch science fiction movies may like "The Three-Body Problem"), similar users may like the same products (user A and user B have similar preferences, user A likes things, maybe user B will also like it), but this is a game of probability.

The system guesses what I most likely like ≠ I really want.

The LLM (Large Language Model) led by ChatGPT can understand human language, and it can better "see" the true intention behind a sentence, that is, it has stronger perception ability, which can just make up for the shortcomings of existing recommendation algorithms.

Today, let's talk about ChatGPT recommendation system vs standard algorithm, what are the advantages? You can also leave a message at the end of the article to discuss, will ChatGPT replace the existing recommendation algorithm?

Putting aside personal opinions, large language models (such as ChatGPT) will impact existing recommendation algorithms, especially after large models are built into smart terminals such as mobile phones, consumers are more inclined to ask questions to get recommendations rather than look at "guess what you like". However, the current mainstream recommendation algorithm may still be embedded in the large model as a recommendation idea.

The advantage of large language models such as ChatGPT is that they can be recommended according to our specific requirements, such as recommending a hotel with a great children's pool, the most Instagram-style restaurant, the quietest vacuum cleaner, etc., which can better meet our individual needs. In contrast, the existing recommendation algorithm recommends by "analogy", which is less attentive.

Of course, ChatGPT also has its drawbacks, and it cannot guarantee 100% accuracy. Moreover, some needs are "tapped". Users may have preferences that they are not aware of and do not communicate with ChatGPT. Collaborative filtering, recommending based on the preferences of similar people, is a more effective idea at present.

Therefore, we are not going to use ChatGPT to "strangle" the existing standard algorithms, but should work hard for better products or technologies.

01 Discover "good fish that slip through the net" outside the algorithm

Existing standard algorithms are often keen to recommend best-selling products:

After all, everyone has certified good things, and there is a high probability that "you" will like them.

ChatGPT's recommendation system is more conducive to discovering neglected "long-tail categories".

Long-tail categories are often difficult to discover and recommend by the original algorithm because they have low sales and lack sufficient user feedback and evaluation. However, these categories may happen to be of interest to some consumers.

ChatGPT can act as this pair of "eyes" and find "good fish" outside the algorithm.

Let's say you buy a book about artificial intelligence, such as A Brief History of Artificial Intelligence.

Under the original recommendation algorithm, the system may continue to recommend other books of the same type, such as "Introduction to Artificial Intelligence", "Artificial Intelligence and Machine Learning", etc. Although these books have similar themes and characteristics to the purchased books, they may have highly overlapping content, such as the origin of artificial intelligence, the application of natural language and other repetitive chapters, which may not necessarily stimulate your interest in reading.

ChatGPT may recommend related but different types of books, such as Artificial Intelligence and Philosophy, Artificial Intelligence and Art, Artificial Intelligence and Society, etc. Although these books are relatively niche, they have different perspectives and content to help you expand your knowledge and horizons.

Therefore, ChatGPT can break through the limitations of existing recommendation algorithms, not only relying on consumer behavior data or product feature data, but also paying attention to the "logical" correlation between products.

In addition, ChatGPT can also make cross-category recommendations, such as recommending relevant courses, lectures, games, products, etc. according to the theme of "artificial intelligence", improving the richness of existing algorithms.

02 From simple "recommendation" to powerful "persuasion"

ChatGPT can make recommendations more reasonable, generate easy-to-understand recommendation reasons that conform to human language habits at the same time as recommendations, so as to achieve from simple "recommendations" to powerful "persuasion", and better help consumers complete decisions.

For example, if we want to buy an economics-related book, we enter the query term "economics" on JD.com, and we will get a full page of the recommended list, such as "Principles of Economics", "The Wealth of Nations", "Fast and Slow Thinking", "Xue Zhaofeng Lecture Notes on Economics" and other hot-selling books. These books are usually labeled with "repeat customers 100,000+" and other popularity-based labels.

When we want to further judge which book is more suitable for our interests and reading level, we need to click on the detail page of each book, view their detailed introduction, table of contents, reviews, etc., and even open other software such as Xiaohongshu to comprehensively compare the advantages and disadvantages of each book before making a decision.

ChatGPT can directly tell us the reason for the choice and help us automatically complete the process of comparison and screening.

For example, ChatGPT might write a recommendation for Principles of Economics like this:

This book is the world's most popular economics textbook, written by Harvard University professor Gregory Mankiw, systematically introduces the basic concepts, principles and methods of microeconomics and macroeconomics, and uses vivid and interesting examples and cases to explain economic phenomena and policies. This book is suitable for readers who have a certain mathematical foundation and want to systematically learn the theory and methods of economics.

For example, ChatGPT might write a recommendation for The Wealth of Nations:

This book is the starting point of modern political economy research, published by the British thinker Adam Smith in 1776, profoundly analyzed free trade, market competition, division of labor, money and banking and other economic issues, put forward "natural order", "invisible hand", "laissez-faire" and other famous theoretical views. This book is suitable for readers who have a certain historical background and want to understand the history of economics and classic ideas.

These testimonials are not only about the book itself, but also about which readers are suitable.

In addition, we can also ask more questions to ChatGPT, such as "Which book is better for beginners?" Which book is more practice-oriented?" "Which book has the highest rating?" ChatGPT can answer accordingly.

03 From "clear needs" to "vague needs"