The multi-model fusion algorithm with multi-model fusion algorithm can be significantly improved than the single model algorithm. But how to effectively integrate and give full play to the strengths of each algorithm? Here is a summary of some common fusion methods: 1. Linear weighted Fusion method linear weighting is the most simple and easy-to-use fusion algorithm, engineering implementation is very convenient, only need to summarize the results of a single model, and then assign different weights according to different algorithms, the results of a number of recommended algorithms are weighted, you can get results:
It is the score of the user's recommended item (item), the weight of the algorithm K, and the user's recommendation score for the item, which is obtained by the algorithm K. This fusion method is simple to implement, but less effective. Because the parameters of linear weighting are fixed, the selection of parameters in practice usually relies on the summary of the global results, and once set, it is not possible to change automatically according to different recommended scenes. For example, if a scene with algorithm a effect is better, another scenario with the algorithm B effect is better, the linear fusion method in this case can not achieve good results. In order to solve this problem, the philosophical data has been improved, by introducing the mechanism of dynamic parameters, by training users to evaluate the recommendation results, and to generate a weighted model with the system's prediction, the dynamic adjustment weights make the effect greatly improved. 2. Cross Fusion is often referred to as the blending method, and the idea is to cross-reference the results of different recommendation models to ensure the diversity of results. This way the results of different algorithms are combined to recommend to the user.
The idea of cross-fusion method is "each flower into the eyes", different algorithm results focus on different, can meet the needs of different users, directly interspersed together to display. This approach applies to recommended scenarios where more than one result is present, and is often used to differentiate between algorithms, such as those obtained based on the user's long-term interest and short-term interest calculations, respectively. 3. Waterfall Fusion Waterfall (Waterfall model) Fusion method uses a method of concatenating multiple models. Each recommended algorithm is treated as a filter, by means of a method of bridging different particle size filters:
In waterfall-type hybrid technology, the result of the previous recommended method filtering, as a candidate for the next recommended method of input, layer by step, the candidate results in this process will be gradually selected, and finally get a small quantity of high quality result set. This design is typically used on recommended scenarios where there are a large number of candidate collections. In the design of waterfall hybrid system, the algorithm with fast operation speed and low sensitivity is usually ranked in the forefront, and gradually transitions into the heavyweight algorithm, so that valuable computing resources are concentrated on the computation of a few higher candidate results. In the face of a large number of candidate recommendation (Item), and the exposure of the recommended results are less, the requirements of high precision, and limited computing time scenarios, often very suitable. 4. The characteristic fusion method has different raw data quality, which has a great influence on the results of the proposed calculation. Taking the user interest model as an example, we can dig out the user's "explicit" interest from the user's actual purchase behavior, and can use the user's click behavior to excavate the user's "implicit" interest; In addition, users ' preferences can be inferred from user classification and demographic analysis. If there is a user's social network, It is also possible to understand the impact of the users around them on the user's interests. So by using different data sources, extracting different features, entering into the recommendation model for training, and then merging the results. This approach solves the problem of missing data that is often encountered in the real world, because not all users have a full range of data, for example, some users lack transactional information, others do not have social relationship data. The method of feature fusion can ensure the model is not picky eaters and enlarge the applicable surface. 5. The predictive Fusion recommendation algorithm can also be seen as a "predictive algorithm" that we predict for each user what he or she is likely to like next. The idea of the fusion method is that we can predict each prediction algorithm again, that is, the prediction results of different algorithms, we can train the second-level prediction algorithm to predict again, and generate the final prediction results. As shown, we use the predictive results of each recommendation algorithm as a feature, the user's feedback data as a training sample, forming a second-tier predictive model training set, the specific process is as follows
The two-layer predictive model can use commonly used classification algorithms, such as SVM, random forest, large entropy, etc., but in the philosophical practice, the fusion effect is GBDT (Gradient boosting decision Tree) method. 6. Classifier boosting thought recommendation problems can sometimes be translated into patterns (pattern classification) to see, we will be the candidate set is recommended to be divided into several different sets, and then through the design of the classifier method to solve. In this way, we can use the boosting idea in the classification algorithm to combine several weak classifiers into a strong classifier. The core idea of boosting is that after each round of training, the sample of prediction error is given a larger weight, and the following training sets are added, that is, the learning algorithm can intensify the difficult case in the follow-up training, so that a predictive function sequence h with weight is obtained, and the predictive function with good prediction effect has a larger weight and vice versa. The final predictive function H uses a weighted voting method to classify the problem, and the new example is judged by the weighting average of the regression problem. The flow of the algorithm is as follows: (reference from Treeboost paper)
The convergence of the model is often better, but the cost of implementation and the computational overhead are also relatively large.
Multi-model Fusion recommendation algorithm
