Chapter Using Contextual Information
5.1 time context information
Information after a given time, the recommendation system into a system when a variation from a static system has a triple (u, i, t) represents a user u i at time t of the article had behavior.
( A) life cycle assessment items indicators:
(1) Average item Online Days: If an item is produced a day in the behavior of at least one user, the item is defined in this day online
(2) spaced apart average similarity degree vector T popularity days article (cosine similarity) to evaluate a system timeliness
Time diversity (b) recommendation algorithm: the recommended daily results of varying degrees of recommendation system is defined as the time diversity recommendation system
(c) In case the user does not act how to ensure that the results have a certain amount of time recommended by diversity:
(1) adding a certain randomness in generating the recommendation result
(2) Record recommendation results users see every day, then recommend right down at the end of the recommended time for the user has already seen heavy
(3) the recommended daily use different algorithms to the user
(D) the time context recommendation algorithm
(1) Recently the most popular: recommendation to the user's most recent most popular items given time T, the popularity of recent articles i $ n_i (T) $ can be defined as:
itemCF algorithm (2) time contextual
The core part itemCF are: the use of user behavior off-line calculation of similarity between the goods; based on user behavior and historical objects similarity matrix, giving users do online personalized recommendations.
Calculated using the time information item to make a similarity matrix improvements:
Equation $ f (\ left | {t_ {u, i} -t_ {u, j}} \ right |) $ attenuation and time-related items, $ t_ {u, i} $ is an article for user u i significance generates the behavior of a function of time is the time the user .f j i generates the behavior of the article and the article spacing greater the smaller the value of f, the following specific functions ($ \ $ Alpha is the attenuation coefficient for the hyper-parameters, need multiple times of training to obtain appropriate values):
Similarly, when the prediction time using information also needs to correct the result of the prediction, the appropriate formula is as follows:
userCF algorithm (3) time-related context
And itemCF similar approach taken, the time decay factor is introduced, not tired out.
(4) FIG period Model
Period graph model G (U, $ S_U $, I, $ S_I $, E, w, $ \ sigma $) is a bipartite graph .U is a combination of user nodes, $ S_U $ user time period set of nodes. A user period node $ V_ {ut} {\ in} S_U $ t will be connected to the user and the article of interest at all times. I is a collection of items, $ S_I $ is a collection of articles periods. A period article node $ V_ {it} {\ in} S_I $ and all users at time t i had a behavior is connected by an article edge. E is the set of edges, mainly comprising three sides (a) user behavior items have u i, there is $ e (v_u, v_i) {\ in} E $; (b) if the user u at time t of the article i acts, then there are two sides $ e (v_ {ut}, v_i), e (v_u, v_ {it}) {\ in} E $. $ w (e) $ defines the weight of an edge, $ \ sigma (e) $ defines a vertex weights. FIG calculating the correlation between the two paths through the vertices of the fusion algorithm.