The reference link of this article is for personal learning only:
https://github.com/datawhalechina/team-learning-rs/tree/master/DeepRecommendationModel
https://blog.csdn.net/qq_xuanshuang/article/details/105730935
Related learning books recommendation: "Deep Learning Recommendation System" Wang Zhe
Description of Wide & Deep
The main function of the Wide part is to make the model have a strong "memory ability"; the main function of the Deep part is to make the model have a "generalization ability".
"Memory Ability": The model's ability to directly learn and use the "co-occurrence frequency" of items or strong features in historical data. Generally speaking, simple models such as collaborative filtering and logistic regression have strong "memory ability". This type of model has a simple structure and does not have too many features. The original data can often directly affect the recommendation results, resulting in a regular recommendation similar to "If you have clicked A, then recommend B", which is equivalent to the model directly remembering historical data And use these memories to make recommendations. The features of the multi-layer neural network will be processed in multiple layers, and will be continuously crossed with other features, making the model's memory of the original strong features inferior to the simple model.
"Generalization ability": The relevance of the model's transfer characteristics, and the ability to detect the relevance of rare features that have never appeared before, and the final label. Matrix factorization is stronger than the generalization ability of collaborative filtering, because MF introduces hidden vectors, so that users or items with sparse data can also generate hidden vectors to obtain recommendation scores supported by data. This is a typical example of passing global data to rare On the items, thereby improving the generalization ability. For another example, a deep neural network can discover potential patterns in the data through feature sub-combination. Even with very sparse feature vector input, it can get a relatively stable and smooth recommendation probability.
Wide & Deep Model
Wide&Deep consists of a shallow (or single-layer) Wide part neural network and a deep Deep part multi-layer neural network. The output layer uses softmax or logistics regression to synthesize the output of the Wide and Deep parts.
Memorization: The
previous large-scale sparse input processing was: through linear model + feature crossover. The memorization and memory ability brought by it are very effective and explainable. But Generalization (generalization ability) requires more manual feature engineering.
Generalization:
In contrast, DNN requires almost no feature engineering. By combining low-latitude dense embedding, deeper hidden features can be learned. However, the disadvantage is that it is a bit over-generalize. The performance of the recommendation system is: it will recommend users that are not so relevant items, especially the user-item matrix is sparse and high-rank (high-rank matrix)
The difference between the two:
Memorization tends to be more conservative, recommending items that users have had behavior before. In contrast, generalization tends to increase the diversity of the recommendation system.
Wide & Deep:
Wide & Deep consists of two parts: linear model + DNN part. Combine the advantages of the above two to balance memoryization and generalization.
Reason: Integrate the advantages of memoryization and generalizatio to serve the recommendation system. Compared with the wide-only and deep-only models, wide & deep has improved significantly.
Wide&Deep model evolution-Deep & Cross model
-
Use Cross Net to replace the Wide part to increase feature crossover strength