Recommended algorithm Learning Note one

"The greatest wealth of the next 30 years will be in this field; the more content expands, the more it needs to be filtered to focus. Where the attention flows, the money goes. --a blog post from a note

During the university because of the conflict (as long as people lazy thinking), not in-depth study of "machine learning" related knowledge. Recently saw a lot about the recommended blog, but also with the cattle students have related aspects of contact, the previous slack made me regret. It is not too late to mend.

But really, aside from the subjective factors, learn to recommend relevant knowledge, to engage in information technology is really beneficial to the industry.

To this end, this article introduces the recommended algorithm (recommended to read the original text directly, link in the last).

# # Algorithm Classification

Recommended algorithms are generally divided into four main categories (1-4):

• Collaborative Filtering recommendation algorithm
• Content-based recommendation algorithm
• Hybrid recommendation algorithm
• Popularity recommendation algorithm

In addition to these kinds, there are some advanced non-traditional recommended algorithms (5).

# # Algorithm Detail Properties

1. Collaborative Filtering recommendation algorithm

• Introduction: Generate user-specific recommendations by looking for certain patterns in a user's series of behaviors

• Input: Relies solely on idiomatic data (such as user preferences such as reviews, purchases, downloads, etc.)

• Type:

  • Neighborhood-based collaborative filtering (user-based and item-based)
  • Model-based collaborative filtering (matrix factor decomposition, restricted Boltzmann machines, Bayesian networks, etc.)

• Advantages:

  • Minimum domain required
  • Users and items are not required
  • In most scenarios, good enough results can be produced.

• Disadvantages:

  • Cold start problem
  • Requires standardized products
  • Requires a high percentage of users and items (1:10)
  • Prevalence bias (not good enough when you have a long tail)
  • Difficult to provide explanations

2. Content-based recommendation algorithms

• Summary: A similar item to a user's recommendation and content (such as metadata, description, topic, and so on) that they have previously liked

• Input: Depends solely on the item and the user's content/description (in addition to idiomatic data)

• Type:

  • Information retrieval (e.g. TF-IDF and Okapi BM25)
  • Machine learning (e.g. naive Bayes, support vector machines, decision trees, etc.)

• Advantages:

  • No cold start problem
  • No usage data required
  • Without prevalence bias, items with rare characteristics can be recommended
  • You can use User Content features to provide explanations

• Disadvantages:

  • Item content must be machine readable and meaningful
  • Easy to archive users
  • It's hard to have surprises, a lack of diversity
  • It is difficult to federate attributes of multiple items

3. Hybrid recommendation algorithm

• Introduction: Comprehensive utilization of collaborative filtering recommendation algorithm and content-based recommendation algorithm each of the advantages of the same time to offset their shortcomings

• Input: Use both user and item content characteristics and idiomatic data while benefiting from two input types

• Type:

  • Weighted
  • Exchange
  • Mixed
  • Feature Combination
  • Case List
  • Feature enhancement
  • Meta-level

• Advantages:

  • By using collaborative filtering recommendation algorithm and content-based recommendation algorithm alone
  • No cold start problem
  • No prevalence bias, can recommend items with rare characteristics
  • Can generate surprises and achieve diversity

• Disadvantages:

  • It takes a lot of work to get the right balance.

4. Popularity recommendation algorithm

• Summary: This is a way to recommend popular items (for example, most downloads, most viewed, and most affected items)

• Input: Use the content of idiomatic data and items (such as category)

• Advantages:

  • Relatively easy to implement
  • A good benchmark algorithm
  • Helps resolve new user cold boot issues

• Disadvantages:

  • Requires standardized products
  • Often require some type of item to classify
  • New items are not recommended (very few opportunities to be observed)
  • Recommendation list does not change too much

5. Advanced non-traditional recommendation algorithm

• Type:

  • Deep learning
  • Learning Level
  • Multi-armed Bandits (Exploration/development)
  • Context-aware recommendations
  • Tensor decomposition
  • Decomposition machine
  • Social recommendation

• Advantages:

  • Helps to barely maintain final performance percentage points
  • You can say you're using a progressive approach

• Disadvantages:

  • Difficult to understand
  • Lack of recommended tool support
  • No recommended way to provide your first referral system

# # Reference

Reference article Links:

http://www.infoq.com/cn/articles/recommendation-algorithm-overview-part01

https://buildingrecommenders.wordpress.com/2015/11/16/overview-of-recommender-algorithms-part-1/

# # Conclusion

"Most importantly, in a world of information-rich, the only scarce resource is human attention." ”

Recommended algorithm Learning Note one