DNN model training method and device

Abstract

The invention discloses a DNN model training method and device, relates to the technical field of machine learning, and can realize joint training of original data sets of companies of two parties onthe premise of not leaking the original data sets of the companies of the two parties. The method comprises the following steps: respectively carrying out unified dimension expansion on dimension vectors in a first data set and a second data set; setting structural parameters of the DNN model; respectively converting data in the first data set and the second data set by using the linear relation coefficient matrix WL and the bias vector bL corresponding to the second hierarchy; performing traversal calculation on each neuron result from the second level to the L level through a loss function and a back propagation function; and correspondingly outputting when the variation of the linear relation coefficient matrix WL and the bias vector bL of each level in the current iteration relative to the last iteration is smaller than an iteration stopping threshold value, otherwise, returning to the data conversion step to re-execute iterative computation until the number of iterations isgreater than the maximum number of iterations.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a DNN model training method and device. Background technique [0002] Enterprises usually have a large amount of data, and each company protects its own data as a core asset to prevent leakage. When companies use their own user data, they usually need to predict whether a user will perform a certain behavior based on the user's label data, for example, According to the user profile, predict whether the user will click on a promotional advertisement. The industry uses various prediction models to predict user behavior, among which DNN is a prediction model with a wide range of applications. [0003] The effectiveness of the DNN model depends on the quality of the training data set, including the quantity and quality of labels for a single user, as well as the number of users. Usually, a single company’s own data is limited, and it is difficult to form a high-quality trai...

AI Technical Summary

Problems solved by technology

[0003] The effectiveness of the DNN model depends on the quality of the training data set, including the quantity and quality of labels for a single user, as well as the number of users. Usually, a single company’s own data is limited, and it is difficult to form a high-quality training set, resulting in The model is not effective enough and the prediction accuracy is low

If the data sets of two companies are used for DNN model training, there are two current methods: the first is to pool data to one of the parties for training, this method will inevitably leak the data of one party to the other party, causing many companies to The scheme is more contradictory; the second is to gather the data of both parties into a neutral third-party environment for training. The problem with this method is that it needs to have enough trust in the third party, trust that it will not leak the data of both parties, and sign a relevant agreement. contract, the process is long, which is not conducive to business promotion

Images