Recommendation System 5th Week---Content-based recommendations, the semantic model LFM

Content-based recommendations

Difficulty in developing classification/attributes
Use Professionals (editors) to sort products, but this creates cost and efficiency bottlenecks
Editorial opinion may not be representative of the user's opinion, subject to the professional level of the editor.
The granularity of classification is difficult to control
If a product has multiple classifications, it is difficult to consider
Multidimensional, multi-angle classification
Editing is hard to determine the weight of a product in a category

The semantic model of the lingo

LFM's former PI This Life

The applicability of the semantic model of the lingo

About training Sets

Solutions to common problems of the same kind

The geometrical meaning of gradient descent method

The extremum of LFM loss function is solved by gradient descent method

The important parameters in LFM

The number of hidden features in a model
The learning rate selected in Gradient descent method
penalty factor Lambda in loss of the number of Taipa
negative sample/positive sample ratio ratio for training set

Movielens Data Set Download

Verifying LFM validity using the Movielens data set

Influence of proportional parameter ratio of positive and negative sample

Several indicators

Advantages and disadvantages of LFM

A typical machine learning algorithm, with a good mathematical basis, looks more mathematical aesthetic
indicators are generally slightly higher than ITEMCF and USERCF
Use less memory during training
The computation time is more than itemcf or USERCF because of the need for iteration
Cannot be calculated online in real time
It is difficult to explain the rationality of the model to the family

Netflix Grand Prix
Netflix, Inc. (NASDAQ:NFLX) is an online video rental provider. The company is able to provide a large number of DVDs, but also allows customers to quickly and easily select the film, and free delivery.

Netflix has been rated as the most satisfying site for customers five times in a row. Movies and e-regulation programs can be viewed via PC, TV, ipad, iphone, and can be connected to TV via WII,XBOX360,PS3 and other devices.
Since October 2006, Netflix has published about 100 million 1-5 anonymous film ratings, with data set kernels containing film titles, rating stars and ratings dates, and no text reviews.

The competition requires contestants to predict what movies Netflix customers like, and to increase the efficiency of the forecast by more than 10%.
Http://baike.baidu.com/view/2836949.htm?fr=aladdin#3
It has a far-reaching influence on the development of recommendation system algorithms, such as the popularity of LFM to quickly enter the public regulations, and a lot of improvement methods for LFM.

Improved LFM
The eighth chapter of Xiangliang Book
The impact of a product's own characteristics (such as quality) by adding a bias to the prediction formula to take into account individual factors (such as the more demanding personality of some reviewers)
Considering the neighborhood effect of LFM, more like itemcf deformation, or the enhanced version of SVD, it is called svd++
Add time variables to the model (taking into account the user's interest will change over time)
Model combinations

Gmt

Detect languageAlbanianArabicAzerbaijani languageIrishEstonianBasque languageBelarusian languageBulgarianIcelandicPolishBosnianPersianBoolean language (Afrikaans)DanishGermanRussianFrenchFilipinoFinnishKhmer languageGeorgian languageGujaratiKazakhHaitian CreoleKoreanHausa languageDutchGalicianCatalanCzechKannada languageCroatianLatin languageLatvianLao languageLithuanianRomanian languageMalagasy languageMalteseMarathiMalayalamMalayFYRO MacedonianMaoriMongolianBengaliBurmese languageHmongZulu, South AfricaNepalese languageNorwegianPunjabiPortugueseChichewa languageJapaneseSwedishSerbian languageSesotho languageSinhala languageWorld languageSlovakSlovenianSwahiliCebu languageSomalia languageTajik languageTeluguTamilThaiTurkishWelshUrdu languageUkrainianUzbek languageHebrewGreekSpanishHungarianArmenianIgbo languageItalianYiddishHindiIndonesian SundaIndonesian languageIndonesian JavaneseEnglishYorubaVietnameseChinese SimplifiedChinese Traditional

AlbanianArabicAzerbaijani languageIrishEstonianBasque languageBelarusian languageBulgarianIcelandicPolishBosnianPersianBoolean language (Afrikaans)DanishGermanRussianFrenchFilipinoFinnishKhmer languageGeorgian languageGujaratiKazakhHaitian CreoleKoreanHausa languageDutchGalicianCatalanCzechKannada languageCroatianLatin languageLatvianLao languageLithuanianRomanian languageMalagasy languageMalteseMarathiMalayalamMalayFYRO MacedonianMaoriMongolianBengaliBurmese languageHmongZulu, South AfricaNepalese languageNorwegianPunjabiPortugueseChichewa languageJapaneseSwedishSerbian languageSesotho languageSinhala languageWorld languageSlovakSlovenianSwahiliCebu languageSomalia languageTajik languageTeluguTamilThaiTurkishWelshUrdu languageUkrainianUzbek languageHebrewGreekSpanishHungarianArmenianIgbo languageItalianYiddishHindiIndonesian SundaIndonesian languageIndonesian JavaneseEnglishYorubaVietnameseChinese SimplifiedChinese Traditional

Language features limited to 100 character options: History: Help: Anti-feedback

Recommendation System 5th Week---Content-based recommendations, the semantic model LFM