Summary of hybrid recommendation technology

Completely reproduced from:A Jun's blog

Hybrid recommendation system is another hot topic of recommendation system research. It refers to mixing multiple recommendation technologies to compensate for their shortcomings, so as to achieve better recommendation results. The most common is to combine collaborative filtering technology with other technologies to overcome the problem of cold start. (1) weighted model is to produce a recommendation by weighting the computing results of multiple recommendation technologies. The simplest method is linear mixing. First, the recommendation results of collaborative filtering and content-based recommendation results are given the same weight, then compare the user's evaluation of the item with the system's prediction, and then adjust the weight value. The weighted hybrid model is characterized by the fact that the entire system performance is directly related to the recommendation process, so that it is easy to adjust the corresponding hybrid model for the trust distribution of watermelon, however, the premise of this technology is that the values of parameters related to different technologies are basically the same for all possible items in the whole space. (2) The conversion model uses different recommendation technologies based on the problem background and actual situation. For example, the system uses Content-based recommendation and collaborative filtering. if the system does not produce highly reliable recommendations, it tries to use collaborative filtering. This method will increase because the conversion standard needs to be compared in various situations. Algorithm Of course, the advantage of doing so is that it is sensitive to the advantages and weaknesses of various recommendation technologies. (3) The merging model uses a variety of Recommendation technologies to provide a variety of Recommendation results, providing users with reference. For example, you can build such a personalized recommendation system based on Web logs and cache data mining. This system first constructs a multi-user interest model by mining Web logs and cache data, then, the system matches the short-term access history of the target user with the user interest model and uses the content-based filtering algorithm to recommend similar webpages to the user. At the same time, it filters out the systems among multiple users, predict the most likely page to be accessed next for the target user, and sort the page according to the score. The page is attached to the current user request access page and then recommended to the user. That is, "You may like webpages that you may be interested in ". (4) feature combination combines features from different recommendation data sources and is adopted by another recommendation technology. Collaborative filtering information is usually used as the added feature vector, and content-based recommendation technology is used for the added dataset. The hybrid feature combination makes the system no longer only consider the data sources for collaborative filtering, so it reduces the user's sensitivity to the number of project ratings. On the contrary, it allows the system to have similar information about items, it is not transparent to the collaborative system. (5) waterfall-based recommendation method optimization the previous recommendation method: it is a phased process. First, a recommendation technology is used to generate a rough candidate result, on this basis, the second recommendation technology is used to make further accurate recommendations. Waterfall allows the system to avoid using low-priority technologies for certain items, which may be well differentiated by the first recommendation technology, or a project that is rarely evaluated by users and never recommended. Because the second step of waterfall is only focused on items that need to be judged separately. In addition, the waterfall type has high fault tolerance in low-priority technology, because the score obtained with high priority will become more accurate, rather than being completely modified. (6) output of the previous recommendation method as the input of the next recommendation method for feature increment. For example, you can use clustering analysis as the preprocessing of association rules. First, you can cluster session files, and then mine association rules for each clustering to obtain association rules for different clusters. After an Access session is obtained, the matching values of the Access session and each cluster are calculated to determine which cluster the Access session belongs to and then the corresponding Association Rules of the cluster are applied for recommendation. What are the differences between this type and waterfall type? In feature incrementing, the features used by the second recommendation method include the first output. In the waterfall model, the second recommendation method does not use the output of any sort generated by the first method. The results of the two recommendation methods are mixed in an optimized way. (7) The meta-hierarchy model uses the model generated by one recommendation method as the input of another recommendation method. This is different from feature incrementing because a learning model is used in feature incrementing to generate certain features as the input of the second algorithm, and the entire model is used as the input in the meta-layered model. For example, you can combine user-based collaborative filtering and project-based collaborative filtering algorithms to first solve similar project sets of the target project, the user-based collaborative filtering algorithm is used in similar project sets of the target project. This collaborative recommendation method for neighboring users based on similar projects can effectively deal with personalized recommendation issues with a large number of users' interests, especially when the content and attributes of candidate recommendation projects differ greatly, this method provides better performance.