Eight, reading papers

Eight, reading papers

• Thesis Title: Xing Chunxiao, high Fengrong, war thinking man, Zhou column adapt to the user's interest changes collaborative filtering recommendation algorithm.
• Studied
  by different research user interest, take the initiative to recommend resources most needed for the user to personalize recommendations.
• Research motivation
  for the timely response to user interest change, proposed two improved (based on weight data right time, and the right resources based on data similarity weight)
• Literature review
  
  • Collaborative filtering algorithms
    typical collaborative filtering algorithm is based on the use of user access to each other's similarities recommended users may be interested in the resource users . Input data recommendation system can be expressed as the user's mxn - resource access matrix R, m is the number of users, n being the number of resources. Matrix values (0, 1) indicates that the user access to resources or not.
    By user (u) the historical record and access specific similarity function, calculate the access behavior of its closest neighbor k users as the most recent set of the user , the statistics of neighbor user has visited, and u unvisited resources to generate recommendations sets, computing resources of the degree of recommendation for each of u, taking the first N as recommended set.

  • Collaborative filtering algorithm based on resource
    collection of resources recommended by the current user has access to the resource is not accessible, the similarity of the user's resources than similar stability, scalability problem solving algorithm.
    Similarity calculation is an important factor affecting the performance of the algorithm is recommended .
    Similarity calculation methods are: cosine similarity, Pearson correlation coefficient, conditional probability and the like.
    The paper uses the conditional probability calculating the similarity between resources, resources for i, j with P | represents (i j) the conditional probability that they can be accessed by the same user to measure the similarity between resources, calculated between i and j similarity formula:

    
    
    \ (sim (i, j) = \ frac {P (i | j)} {Freq (i) ^ {\ alpha}} = \ frac {Freq (i, j)} {Freq (j) \ times Freq ( I) ^ {\ Alpha}} \)
    
    Freq is the number of users accessing a resource, [alpha] is a number between 0 and 1, called object scaling factor [alpha] is to weaken the introduction of the resource being accessed many times similarity the impact of computing.
  • Data token time weighting function based on
    
    
    $ the WT (U, i) = (l- \ Alpha) + \ Alpha \ FRAC {D_ {UI}} {L_ {U}} $
    
    
    the WT (U, i) represents the resource i of the user weight, D_ {ui} is the time the user access to the resource i of the user u first access a resource time intervals, L_ {u} is the time recommendation system span, a ∈ (0, 1) weighted exponential growth, can be dynamically adjusted .
    
    
  • Based on the resource similarity weighting function
    
    
    \ (WS (u, i) = \ overline {sim (i, I_ {uT})} = \ frac {\ sum_ {j \ in I_ {uT}} ^ {} sim (i, J) {size} (I_ {uT})} \)
    
    
    size (I_ {uT}) I_ {uT} represents the number of resources calculated WS (u, i) to be calculated similarity and I_ {uT} i for each resource degree
    
    
  • Two kinds of weights binding
    
    
    \ (WTS (u, i)
    
    
    = \ beta \ times WT (u, i) + (1- \ beta) \ times WS (u, i) \) scale factor β∈ [0,1], β (1 -β) represents both of the weight and weight percentage values obtained ratio.
    
    
• Study Design
  Design of the experiment Group 3, comparative tests.
• Use the data set
  KDD2000 online transaction data sets
• Analysis conclusion
  
• Learning experience
  parameters to be selected according to different schemes in different recommendation system, the paper last-mentioned weight function automatically determines the parameters of the right, is a good entry point for research.