Introduction to the explicit factor model

Sigir contains three valuable papers in the recommendation system, and proposes a new algorithm framework. This article introduces the first algorithm framework (from thesis: Explicit Factor Models for explainable recommendation based on phrase-level sentiment analysis, An interpretive recommendation model based on phrase-level affective analysis--explicit factor model. If you have different understanding of this article, do not hesitate to enlighten.

I. OverviewEFM (Explicit Factor Models, explicit factor model) is designed for the shortcomings of LFM (latent Factor Models, Hidden factor model). the features of LFM are as follows:A. Abstract the hidden factor space by classification. In the classification process, we do not need to care about the angle of classification, the results are based on user scoring automatic clustering.       The granularity of the classification is controlled by setting the final classification number of LFM.       B. For each item, it is not explicitly divided into a category, but rather the degree to which it belongs. C. For each user, calculate his interest in each class. The shortage is:<1> A single rating does not reflect the user's preferences for the characteristics of the item, and does not take advantage of user reviews. <2> because the categories are abstract and have no clear meaning, it is not possible to explain the recommended reasons when recommending items to users. the characteristics of EFM are as follows:A. By phrase-level (phrase-level) sentiment analysis of user comments, the characteristics of the item and the user's opinion are extracted explicitly.       B. For each item, calculate how much it contains for each feature.       C. For each user, calculate how much he likes each feature.       D. The user-item preference matrix is calculated based on the user's comments and scoring data (setting the weights for both). E. When recommending purchases to users, users are advised not to purchase certain items. The advantage is:<1> make full use of user comments to improve the accuracy of the algorithm. <2> because the characteristics of the goods have been extracted explicitly, so to recommend products to users, you can intuitively explain the reasons for the recommendation. This helps users to decide whether to buy or not, especially if they are advised not to purchase certain items, which can help increase user confidence in the system. second, EFM algorithm framework 1. Build an emotional DictionaryEFM the process of building a dictionary is illustrated by the following example: ( ShadowsThe grid indicates that the user commented on the item. First, from the user's comment CorpusExtract the characteristics of an item (or, on one aspect of an item): Screen, earphone. The user's views on these features are then extracted: perfect, good. If these expressions of opinion The vocabulary itselfIt is positive emotion, then it is expressed in 1, whereas the other is 1.      So in this case, the emotional phrase is expressed as (screen, perfect, 1), (earphone, good, 1), and this phrase constitutes an emotional dictionary. Emotional analysis of user reviews based on affective dictionaries, determine whether the user's feelings are positive or negative。 For example: Perfect is positive, and good is negative, because the negative word is not in front. So, in this example, the user's comments can be expressed as Features/EmotionsYes: (screen, 1), (earphone,-1). The user's comments are expressed as Features/EmotionsYes, the purpose of building an emotional dictionary. 2. Build The MatrixEFM needs to build three matrices. The first one is user scoring matrix a, which indicates the score of the first user of article J. Because the user does not necessarily have to score all items, so no rating is recorded as 0。 The second one is user-feature focus matrix X, which indicates the degree of preference of the first user to the J feature: where N represents the highest score (typically 5 points) for the user rating.       In order to make the value range of each value of the matrix and the user score matrix is [1, N], the value of the parameter is normalized with the sigmoid function. The third one is Item-Feature mass matrix y, indicating the degree to which the article I contains the J feature: The K indicates that the J feature of the article I was mentioned several times by the user. The K-Times refers to the K-trait/sentiment pair, which calculates the mean value of the K pairs (1 or-1). 3. Estimating missing values for matrices X, Y, aThe non-zero in matrix X, Y represents the relationship between an existing user or an item and a feature, and 0It is not clear that the Missing Value。 To estimate these missing values, use the Optimization loss functionThe method. loss Functionis to put an event MappingTo A real number that can represent the cost of economic or opportunity associated with itOne of the functions. In statistics, loss functions are often used to Estimation Parameters。 For unknown parameters of the loss function θIndicates that the decision-making scheme (the actual value obtained) is used DIndicates that there are two common loss functions: two loss function: L (θ,d) = C (θ−d) 2Absolute loss function: The algorithm uses the Two-time loss function。 The method of using the optimization loss function is to refer to minimizing the difference between the estimated value and the real value。 So the optimal loss function for x and Y is as follows: EFM has extracted explicit features compared to LFM. We assume that some characteristics belong to a certain type, and the user likes this type or the item contains this type, because the characteristic is explicit, thus introduces the concept of "explicit factor". In the expression above R is the number of explicit factors。 Similarly, it is estimated that the missing value of the scoring matrix A also uses the explicit factor. At the same time, taking into account users also take into account other potential factors when scoring, so it also introduces the use of LFM Hidden Factors， with indicates the number of hidden factors。 The optimal loss function for A is:, and then combine the two loss functions into: ( * )Among them, is to prevent over-fitting of the regularization term. ( * )The formula for updating the Matrix V, U1, U2, H1, H2 is obtained by the Lagrange function and the KKT condition, as follows: Set the iteration number to iterate, or after the parameter converges, get the parameter values of the above 5 matrices, thus estimating the missing values of X, Y, a:,   4. Top-k RecommendationsThe line of the vector represents the degree of preference of the user for each feature, selected Subscript for K-Features with maximum parameter value, with the expression. The first user is then calculated by using the following method to score the J item: where, the specific value is determined by the experiment.       In most scoring systems, the highest score is 5, so n=5. Finally, choose the highest rated top K items to recommend to the user, and according to the characteristics of the user to explain the recommendation reasons.

Introduction to the explicit factor model