Method and apparatus for predicting test scores

A method for predicting a test score of a user through an artificial intelligence (AI) model by a terminal, the method comprising:

• generating the AI model by using a pre-training model in which pre-training has been completed, wherein the AI model includes an embedding layer, a core network, and a prediction layer, and the core network includes a compressive transformer configured to process time series data through a compression function, wherein the generating of the AI model includes tuning the pre-training model with all parameters of layers of the pre-training model except a prediction layer of the pre-training model being fixed;
• delivering training data of the user to the embedding layer for embedding;
• embedding the training data through the embedding layer;
• delivering an embedding vector from the embedding layer to the core network including the compressive transformer;
• delivering an output value from the core network to the prediction layer for predicting the test score;
• and
• outputting a prediction value for predicting the test score from the prediction layer,
• wherein the training data are configured with a set of pairs containing questions and whether the user answered the questions correctly,
• wherein the embedding vector is created based on a following equation:
• xn=Eq(qn)+Ea(an-1), and
• wherein the xn means an n-th question, wherein the qn means an n-th embedding vector, the and the an-1 means the user answered a (n−1)-th question correctly, the Eq means an embedding layer related to a question, and Ea means an embedding layer related to whether the user answered the question correctly,
• wherein the delivering of the output value from the core network to the prediction layer comprises:
• inputting an input sequence corresponding to the embedding vector to the core network,
• dividing the input sequence by a unit length set to the compressive transformer,
• sequentially processing fragments of the divided input sequence, starting from a first fragment, by using the compressive transformer,
• based on a new fragment being allocated to the compressive transformer, moving a previously processed fragment to a memory, and
• based on the memory being saturated, compressing fragments from an oldest fragment in the memory according to a specific ratio and moving the compressed fragments to a compressed memory.