For this particular tutorial, we will build a Closed-Form Multiple Linear Regression algorithm and use it to forecast AAVE's (WETH Pool) future projected Lifetime Repayments as a practical example. Towards the second half-end of the tutorial, we will convert the model to Cairo enabling us to make the entire MLR system as well as the forecasts fully provable & verifiable.
Provability & Verifiability
The key benefit of this Lightweight Multiple Linear Regression Solver lies in its commitment to Provability and Verifiability. By utilizing Cairo & Orion, the entire MLR system becomes inherently provable through STARKs, offering unparalleled transparency. This enables every inference of the model construction, execution, and prediction phase to be transparently proved using e.g. LambdaClass STARK Prover. In essence, the Provability and Verifiability aspect ensures that the tool is not only for prediction but also a framework to build accountability and trust in on-chain business environments.
Brief intro to MLR
To give a brief overview of MLR, it is used to model the relationship between a single dependent variable denoted as y, and multiple independent variables, such as x1, x2, etc. This method extends the principles of simple linear regression by allowing us to incorporate multiple explanatory factors to predict y. The significant advantage lies in its capability to evaluate both individual and joint linear relationships between each feature and the target variable, providing a comprehensive understanding of how changes in predictors correspond to changes in the outcome.
The regression coefficients, illustrated by β0, β1...βn play a pivotal role in quantifying the impact of each feature variable on the dependent variable. It not only enables us to discern the individual impact (magnitude and direction) but also unveils how they collaboratively combine to shape outcomes.
In summary, when incorporating multiple factors into our model, we can improve the prediction & forecasting accuracy when compared to relying solely on a single predictor, as seen with simple regression. This enhancement can be mainly attributed to the fact that real-world outcomes being typically influenced by a myriad of factors. Therefore, leveraging multiple linear regression (MLR) serves as a foundational stepping stone to adeptly capture the intricate relationships between features and labels, ultimately guiding us in building accurate and highly interpretable models.
Closed-form approach for computing MLR gradients
As outlined above, MLR still remains a powerful tool for problem-solving in many data-oriented business applications. As we step into the ProvableML domain to enhance model transparency, these algorithms still prove to be highly advantageous in on-chain environments due to their lightweight, interpretable, and cost-efficient attributes.
Traditionally, the common approach to MLR involves computing pseudo-inverses and Singular Value Decomposition (SVD). While robust, their implementation complexity can often overshadow the regression problem at hand. Consequently, gradient-based methods are often preferred in data science projects, but this also can be deemed excessive due to the resource-intensive iterative approach taken to approximate gradients which can be very costly. In addition to this, the manual hyperparameter tuning required can be a significant hindrance, especially in automated on-chain environments.
In light of these considerations, this tutorial introduces an intuitive closed-form approach to calculating MLR gradients without any hyperparameter tuning, making it easy to implement and run MLR algorithms effectively on Starknet. This approach also makes it easy to estimate computational steps/costs required to run MLR given a dataset.
The closed-form MLR comprises of three integral components:
Orthogonalization of Input Features: Ensures independence among the X features.
Gradient Calculation: Computes the exact gradient between each decorrelated X feature and y variable.
Forecasting & Predictions: Utilizes the computed coefficients to make new predictions.
Python implementation
To demonstrate a realistic end-to-end implementation, we'll first work with the AAVE dataset before delving into the implementation of the MLR Solver. Step by step, we'll implement the full process in Python first, which should lay the groundwork to allow us to make a seamless transition to Cairo in the subsequent stages of this tutorial.
Preparing the AAVE dataset
To begin with, we will use the Aave dataset which can be accessed from this link. We will work with our cleaned-up version of the dataset which includes various business metrics such as liquidity incentives and borrowing rates, providing valuable insights for forecasting future lifetime repayments.
import pandas as pdimport numpy as npimport matplotlib.pyplot as pltimport osfrom sklearn.metrics import r2_score# dataset pulled from https://app.aavescan.com/ df_main= pd.read_csv('AAVE-V3-weth.csv')df_main.drop('Unnamed: 0', axis=1, inplace=True)# Order the DataFrame from the oldest to the most recent datapoint based on the datedf_main= df_main.iloc[::-1]#Since Most of the df values are in wei we divide all values by a fixed factor to make the data easy to work with.# Dividing by 1e+22 converts values to thousands of ETH and prevents overflow as we transition to Cairo later.factor =1e+22df_main = df_main/factordays_to_forecast =-7# Our y variable to train ondf['lifetimeRepayments_7day_forecast']= df[['lifetimeRepayments']].shift(days_to_forecast)df = df[0:days_to_forecast]
accruedToTreasury
availableLiquidity
lifetimeFlashLoans
lifetimeLiquidity
lifetimeReserveFactorAccrued
totalLiquidityAsCollateral
totalScaledVariableDebt
lifetimeRepayments
variableBorrowRate
lifetimeRepayments_7day_forecast
30
0.000933
7.40
12.4
204.0
0.0550
119.0
28.7
76.4
3370.0
77.4
29
0.001340
7.45
12.4
205.0
0.0550
119.0
28.7
76.4
3370.0
77.4
28
0.001740
8.02
12.4
207.0
0.0550
119.0
28.8
76.5
3320.0
77.5
27
0.002120
9.59
12.8
212.0
0.0550
124.0
28.6
77.0
3180.0
77.6
26
0.000017
10.00
14.1
215.0
0.0575
128.0
28.7
77.3
3150.0
78.0
In order to separate the feature and label of our dataset, we have replicated the lifetime repayments column into a new target variable column whilst shifting its values up by 7 rows. This aligns each repayment value with the appropriate features from 7 days prior. Consequently, the lifetimeRepayments_7day_forecast column will serve as our predictive label (y), while the other metrics across the same rows become our explanatory variables (X) for predicting future repayments.
By framing our features and labels in this format, we will be able to train the MLR model to be able to estimate the daily lifetime repayments based on current lending pool metrics.
Data normalization
We will now normalize the data using min-max scaling to transform all features and labels into a common 0-1 range.
defnormalize_data(original_data): data_min = np.min(original_data, axis=0) data_max = np.max(original_data, axis=0) data_range = data_max - data_min data_normalized = (original_data - data_min) / data_rangereturn data_normalized#Drop the y label from dataframefeatures = df.drop(['lifetimeRepayments_7day_forecast'], axis=1)#setting our y labeltarget = df['lifetimeRepayments_7day_forecast']# convert data to numpy formatX_original = features.to_numpy()Y_original = target.to_numpy()# normalize the data X_normalized=normalize_data(X_original)y_normalized=normalize_data(Y_original)
Computing MLR gradients
As outlined in the prior section, this closed-form approach to computing the regression coefficients does not rely on gradient descent. Instead, it orthogonalizes the x feature variables, ensuring independence across predictors. It then calculates the gradient between the orthogonalized x features and the y variable. This approach allows us to compute the exact coefficients in a single step, eliminating the need for iterative approximations.
It's very important to notice that in the decorrelate_features function, only the last feature row is fully orthogonalized. The rest of the features are decorelated from one another but are not fully orthogonal to each other. This is done to save on computational costs and make the algorithm more efficient since we can still compute the coefficients without necessarily needing to fully orthogonalize them.
This is better illustrated in the calculate_gradients function, as the process starts from the last fully orthogonalized X feature. Subsequently, it then computes the corresponding gradient and removes this feature's influence from the y label. By iteratively repeating this process across all features we can compute the gradient without the need to have all features fully orthogonalized since we are also removing their influences from the y label iteratively. This streamlined approach reduces computational steps and memory requirements, enhancing the algorithm's efficiency and performance.
#We will first transpose the X features and add a bias term.deftranspose_and_add_bias(feature_data): transposed_data= feature_data.T transposed_data_with_bias = np.vstack((transposed_data, np.ones(transposed_data.shape[1])))return transposed_data_with_bias# decorrelate the features (only the last feature row will be fully orthogonal)defdecorrelate_features(feature_data): x_temp = feature_data.copy() feature_rows = feature_data.shape[0]# Decorrelate featuresfor i inrange(feature_rows): feature_squared = np.sum(x_temp[i]**2)for j inrange(i+1, feature_rows): feature_cross_prod = np.sum(x_temp[i] * x_temp[j])if feature_squared ==0:print('Warning, division by zero encountered and handled') feature_squared =1e-8 feature_grad = feature_cross_prod / feature_squared x_temp[j]-= feature_grad * x_temp[i] decorelated_x_vals = x_tempreturn decorelated_x_vals# compute the exact gradients for each feature variable, including the bias termdefcalculate_gradients(decorelated_x_vals,y_values,original_x_features): y_temp = y_values.copy() feature_rows = decorelated_x_vals.shape[0] gradients = np.zeros(feature_rows)# Calculate gradientsfor i inrange(feature_rows-1, -1, -1): prod = np.sum(y_temp * decorelated_x_vals[i]) squared = np.sum(decorelated_x_vals[i] * decorelated_x_vals[i])if squared ==0:print('Warning, division by zero encountered and handled') squared =1e-8 gradients[i]= prod / squared y_temp -= gradients[i]* original_x_features[i]return gradientsX_normalized_transposed_with_bias =transpose_and_add_bias(X_normalized)decorrelated_X_features =decorrelate_features(X_normalized_transposed_with_bias)gradient_values =calculate_gradients(decorrelated_X_features, y_normalized, X_normalized_transposed_with_bias )real_gradient_values_reversed = np.flip(gradient_values)print('All regression coefficient values, including the bias term: ', real_gradient_values_reversed )>> All regression coefficient values, including the bias term: [-1.27062243 1.15931271 0.173401 -0.31112069, 1.09338439 0.93959362 -1.12956438 -0.08371113 1.18734043 0.3425375 ]
Reconstructing the y labels using the calculated gradients and X feature data
Using the computed regression coefficients we can now rebuild the y labels to see how well they fit to the dataset. In order to achieve this we simply compute the dot product between the calculated coefficient values and original X feature values.
Forecasting the upcoming 7-Day Lifetime Repayments Projections for AAVE's WETH Pool
With the model now fitted, we can use the most recent data points to forecast future repayments projections. Additionally, we will calculate the uncertainty bounds of a 95% confidence interval for these predictions to quantify the reliability of our repayment projections based on the model's historical accuracy across the training data. By using both estimates of prediction and confidence intervals, we provide both repayment expectations and precision guidance that can help in business planning.
Now that we have covered all the steps for constructing and fitting the MLR model using the AAVE dataset in Python, our subsequent step will be to implement it in Cairo. This transition will provide end-to-end provability across all aspects of the multiple linear regression system.
In order to catalyze our development we will leverage Orion's built-in functions and operators to construct the MLR Solver and use it to forecast AAVE's Lifetime Repayments.
Code Structure
The outlined code structure below should serve as a guide to help with our implementation as we will be working within multiple folders.
Scarb is the Cairo package manager specifically created to streamline our Cairo development process. Scarb will typically manage project dependencies, the compilation process (both pure Cairo and Starknet contracts), and downloading and building external libraries such as Orion.You can find all the information about Scarb and Cairo installation here.
To create a new Scarb project, open your terminal and run:
scarbnewmultiple_linear_regresion
A new project folder should be created for you and make sure to replace the content in Scarb.toml file with the following code:
Now let's replace the contents of src/lib.cairo with the following code. This will let our compiler know which files to include during the compilation of our code.
To convert the AAVE dataset to Cairo let's execute the following Python code. This simply creates a new datasets folder for us and converts the x and y variables into Orion's 16x16 tensor format.
Orion's 16x16 tensor format was chosen for this particular tutorial, due to having a relatively good degree of accuracy for both the integer part and decimal part relative to our AAVE dataset.
The converted x and y values will now be populated into aave_x_features.cairo and aave_y_labels.cairo, which should be found under the src/dataset/aave_data folder.
On the other hand, the aave_weth_revenue_data_input will populated into src/dataset/user_inputs_data which is a separate folder. The aave_weth_revenue_data_input represents the latest AAVE's WETH lending pool metrics, which will be later used for performing the 7-day lifetime repayments forecasts.
Now that we have placed the files into this new folder structure, we need to make sure that the files are still accessible to the compiler. Hence, let's create the files aave_data.cairo and user_inputs_data.cairo and add the following module references accordingly.
// in aave_data.cairomod aave_x_features;mod aave_y_labels;
// in user_inputs_data.cairomod aave_weth_revenue_data_input;
Data Preprocessing
Now that our dataset has been generated, it is crucial to implement data normalization before feeding it into the MLR Solver. This is highly recommended for any future MLR implementation in Cairo to mitigate potential overflow issues during subsequent stages. This is due to the MLR closed-form approach involving squaring x values, which can get very large if left unnormalized.
To facilitate this process, we will establish a dedicated Cairo file named data_preprocessing.cairo which should be located under the main src folder. This file will store all our data preprocessing functions, including the min-max normalization function.
// importing libsuse orion::operators::tensor::{Tensor, TensorTrait, FP16x16Tensor, U32Tensor, U32TensorAdd, FP16x16TensorSub, FP16x16TensorAdd,FP16x16TensorDiv, FP16x16TensorMul};use orion::numbers::{FP16x16, FixedTrait};use multiple_linear_regresion::helper_functions::{ get_tensor_data_by_row, transpose_tensor, calculate_mean, calculate_r_score, normalize_user_x_inputs, rescale_predictions};#[derive(Copy, Drop)]structDataset { x_values:Tensor<FP16x16>, y_values:Tensor<FP16x16>,}#[generate_trait]implDataPreprocessing of DatasetTrait {fnnormalize_dataset(ref self:Dataset) ->Dataset {letmut x_values =TensorTrait::<FP16x16>::new(array![1].span(), array![FixedTrait::new(0, false)].span());letmut y_values =TensorTrait::<FP16x16>::new(array![1].span(), array![FixedTrait::new(0, false)].span());// used for multiple_linear_regression_modelsif self.x_values.shape.len() > 1 { x_values =normalize_feature_data(self.x_values); y_values =normalize_label_data(self.y_values); }// used for linear_regression_modelsif self.x_values.shape.len() ==1 { x_values =normalize_label_data(self.x_values); y_values =normalize_label_data(self.y_values); }returnDataset { x_values, y_values }; }}// normalizes 2D Tensorfnnormalize_feature_data(tensor_data:Tensor<FP16x16>) ->Tensor<FP16x16> {letmut x_min_array =ArrayTrait::<FP16x16>::new();letmut x_max_array =ArrayTrait::<FP16x16>::new();letmut x_range_array =ArrayTrait::<FP16x16>::new();letmut normalized_array =ArrayTrait::<FP16x16>::new();// transpose to change rows to be columnslet transposed_tensor = tensor_data.transpose(axes:array![1, 0].span());let tensor_shape = transposed_tensor.shape;let tensor_row_len =*tensor_shape.at(0); // 13 let tensor_column_len =*tensor_shape.at(1); //50// loop and append max and min row values to the corresponding arrayletmut i:u32=0;loop {if i >= tensor_row_len {break (); }letmut transposed_tensor_row =get_tensor_data_by_row(transposed_tensor, i); x_max_array.append(transposed_tensor_row.max_in_tensor()); x_min_array.append(transposed_tensor_row.min_in_tensor()); x_range_array.append(transposed_tensor_row.max_in_tensor() - transposed_tensor_row.min_in_tensor()); i +=1; };// convert array to tensor format for ease of math operationletmut x_min =TensorTrait::<FP16x16 >::new(shape:array![1, tensor_row_len].span(), data: x_min_array.span());letmut x_range =TensorTrait::<FP16x16 >::new(shape:array![1, tensor_row_len].span(), data: x_range_array.span());let normalized_tensor = (tensor_data - x_min) / x_range;return normalized_tensor;}// normalizes 1D tensorfnnormalize_label_data(tensor_data:Tensor<FP16x16>) ->Tensor<FP16x16> {letmut tensor_data_ = tensor_data;letmut normalized_array =ArrayTrait::<FP16x16>::new();letmut range = tensor_data.max_in_tensor() - tensor_data.min_in_tensor();// loop through tensor values normalizing and appending to a new arrayletmut i:u32=0;loop {match tensor_data_.data.pop_front() {Option::Some(tensor_val) => {letmut diff =*tensor_val - tensor_data.min_in_tensor(); normalized_array.append(diff / range); i +=1; },Option::None=> { break; } }; };// convert normalized array values to tensor formatletmut normalized_tensor =TensorTrait::<FP16x16 >::new(shape:array![tensor_data.data.len()].span(), data: normalized_array.span());return normalized_tensor;}
Looking at the code above, we also have implemented a new Dataset struct to encapsulate the predictor features (x_values) and target variable (y_values) into a single reusable data object. By bundling x and y into Dataset, we can easily implement new methods into it such as the normalize_dataset(), allowing for a seamless normalization of both components simultaneously. This approach not only streamlines normalization operations in a single step but also eliminates redundant logic.
The MLR Solver in Cairo
To keep everything organized let's now make a new folder named model under the main src folder. Within it, we will create a dedicated Cairo file named multiple_linear_regression_model.cairo to host all our MLR functions in Cairo.
All of the function MLR functions implemented can be seen below:
use orion::operators::tensor::{Tensor, TensorTrait, FP16x16Tensor, U32Tensor, U32TensorAdd, FP16x16TensorSub, FP16x16TensorAdd,FP16x16TensorDiv, FP16x16TensorMul};use orion::numbers::{FP16x16, FixedTrait};use multiple_linear_regresion::data_preprocessing::{Dataset, DatasetTrait};use multiple_linear_regresion::helper_functions::{ get_tensor_data_by_row, transpose_tensor, calculate_mean, calculate_r_score, normalize_user_x_inputs, rescale_predictions};#[derive(Copy, Drop)]structMultipleLinearRegressionModel { coefficients:Tensor<FP16x16>}#[generate_trait]implRegressionOperation of MultipleLinearRegressionModelTrait {// reconstruct the y values using the computed gradients and x valuesfnpredict(ref self:MultipleLinearRegressionModel, feature_inputs:Tensor<FP16x16> ) ->Tensor<FP16x16> {// random tensor value that we will replaceletmut prediction_result =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![FixedTrait::new(10, false)].span());letmut result =ArrayTrait::<FP16x16>::new();// for multiple predictionsif feature_inputs.shape.len() > 1 {let feature_values =add_bias_term(feature_inputs, 1);letmut data_len:u32=*feature_values.shape.at(0);letmut i:u32=0;loop {if i >= data_len {break (); }let feature_row_values =get_tensor_data_by_row(feature_values, i);let predicted_values = feature_row_values.matmul(@self.coefficients); result.append(*predicted_values.data.at(0)); i +=1; }; prediction_result =TensorTrait::<FP16x16 >::new(shape:array![result.len()].span(), data: result.span()); }// for single predictions if feature_inputs.shape.len() ==1&& self.coefficients.data.len() > 1 {let feature_values =add_bias_term(feature_inputs, 1); prediction_result = feature_values.matmul(@self.coefficients); }return prediction_result; }}fnMultipleLinearRegression(dataset:Dataset) ->MultipleLinearRegressionModel {let x_values_tranposed =transpose_tensor(dataset.x_values);let x_values_tranposed_with_bias =add_bias_term(x_values_tranposed, 0);let decorrelated_x_features =decorrelate_x_features(x_values_tranposed_with_bias);let coefficients =compute_gradients( decorrelated_x_features, dataset.y_values, x_values_tranposed_with_bias );returnMultipleLinearRegressionModel { coefficients };}//Adds bias term to features based on axisfnadd_bias_term(x_feature:Tensor<FP16x16>, axis:u32) ->Tensor<FP16x16> {letmut x_feature_ = x_feature;letmut tensor_with_bias =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![FixedTrait::new(10, false)].span());letmut result =ArrayTrait::<FP16x16>::new();// check if feature data has multiple rows and columnsif x_feature.shape.len() > 1 {letmut index:u32=0;if axis ==1 { index =0; } else { index =1; }let data_len =*x_feature.shape.at(index); // 50letmut i:u32=0;loop {if i >= data_len {break (); } result.append(FixedTrait::new(65536, false)); //65536=ONE in FP16x16, change accordingly i +=1; };if axis ==0 {let res_tensor =TensorTrait::new( shape:array![1, data_len].span(), data: result.span() ); tensor_with_bias =TensorTrait::concat(tensors:array![x_feature, res_tensor].span(), axis: axis); } else {let res_tensor =TensorTrait::new( shape:array![data_len, 1].span(), data: result.span() ); tensor_with_bias =TensorTrait::concat(tensors:array![x_feature, res_tensor].span(), axis: axis); } }// check if feature data is 1Dif x_feature.shape.len() ==1 {letmut j:u32=0;loop {match x_feature_.data.pop_front() {Option::Some(x_val) => { result.append(*x_val); j +=1; },Option::None=> { break; } }; }; result.append(FixedTrait::new(65536, false)); //65536=ONE in FP16x16, change accordingly tensor_with_bias =TensorTrait::<FP16x16>::new(shape:array![result.len()].span(), data: result.span()); }return tensor_with_bias;}// decorrelates the feature data (*only the last tensor row of the decorrelated feature data will be fully orthogonal)fndecorrelate_x_features(x_feature_data:Tensor<FP16x16>) ->Tensor<FP16x16> {letmut input_tensor = x_feature_data;letmut i:u32=0;loop {if i >=*x_feature_data.shape.at(0) {break (); }letmut placeholder =ArrayTrait::<FP16x16>::new();letmut feature_row_values =get_tensor_data_by_row(input_tensor, i);letmut feature_squared = feature_row_values.matmul(@feature_row_values);// avoiding division by zero errorsif*feature_squared.data.at(0) ==FixedTrait::new(0, false) { feature_squared =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![FixedTrait::new(10, false)].span()); }// loop through remaining tensor data and remove the individual tensor factors from one another letmut j:u32= i +1;loop {if j >=*x_feature_data.shape.at(0) {break (); }letmut remaining_tensor_values =get_tensor_data_by_row(input_tensor, j);let feature_cross_product = feature_row_values.matmul(@remaining_tensor_values);let feature_gradients = feature_cross_product / feature_squared; remaining_tensor_values = remaining_tensor_values- (feature_row_values* feature_gradients); //remove the feature factors from one another // loop and append the modified remaining_tensor_values (after the correlated factor has been removed) to the placeholder array
letmut k:u32=0;loop {if k >= remaining_tensor_values.data.len() {break (); } placeholder.append(*remaining_tensor_values.data.at(k)); k +=1; }; j +=1; }; // convert placeholder array to tensor format and update the original tensor with the new modified decorrelated tensor row values
letmut decorrelated_tensor =TensorTrait::new( shape:array![*x_feature_data.shape.at(0) -1- i, *x_feature_data.shape.at(1)].span(), data: placeholder.span() );letmut original_tensor = input_tensor.slice( starts:array![0, 0].span(), ends:array![i +1, *x_feature_data.shape.at(1)].span(), axes: Option::None(()), steps:Option::Some(array![1, 1].span()) ); input_tensor =TensorTrait::concat( tensors:array![original_tensor, decorrelated_tensor].span(), axis:0 ); i +=1; };return input_tensor;}// computes the corresponding MLR gradient using decorrelated featurefncompute_gradients( decorrelated_x_features:Tensor<FP16x16>, y_values:Tensor<FP16x16>, original_x_tensor_values:Tensor<FP16x16>) ->Tensor<FP16x16> {letmut gradient_values_flipped =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![FixedTrait::new(10, false)].span());letmut result =ArrayTrait::<FP16x16>::new();letmut tensor_y_vals = y_values;letmut i:u32=*decorrelated_x_features.shape.at(0);// loop through Decorrelated_x_features starting from the fully orthogonalized last tensor row valuesloop {if i <=0 {break (); }let index_val = i -1;letmut decorelated_feature_row_values =get_tensor_data_by_row( decorrelated_x_features, index_val ); letmut decorelated_features_squared = decorelated_feature_row_values.matmul(@decorelated_feature_row_values);letmut feature_label_cross_product = tensor_y_vals.matmul(@decorelated_feature_row_values); // multiply the tensors// avoiding division by zero errorsif*decorelated_features_squared.data.at(0) ==FixedTrait::new(0, false) { decorelated_features_squared =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![FixedTrait::new(10, false)].span()); } // computing the feature gradient values using the y values and decorrelated x features and appending them to array
letmut single_gradient_value = feature_label_cross_product/ decorelated_features_squared; // divide the summed value by each other result.append(*single_gradient_value.data.at(0));// remove the associated feature gradient value away from y valuesletmut original_x_tensor_row_values =get_tensor_data_by_row( original_x_tensor_values, index_val ); tensor_y_vals = tensor_y_vals- (original_x_tensor_row_values* single_gradient_value); //remove the first feature from the second feature values i -=1; };// convert the gradient array to tensor formatlet final_gradients =TensorTrait::new( shape:array![*decorrelated_x_features.shape.at(0)].span(), data: result.span() );letmut reverse_grad_array =ArrayTrait::<FP16x16>::new();letmut data_len:u32= final_gradients.data.len();loop {if data_len <=0 {break (); }let temp_val = data_len -1; reverse_grad_array.append(*final_gradients.data.at(temp_val)); data_len -=1; };// convert gradient values to tensor formatlet gradient_values_flipped =TensorTrait::<FP16x16 >::new(shape:array![reverse_grad_array.len()].span(), data: reverse_grad_array.span());return gradient_values_flipped;}
At the core of this file lies the pivotal MultipleLinearRegression() function, which orchestrates the entire model fitting process. This function plays a central role by invoking critical functions such as Decorrelate_x_features(), add_bias_term(), and compute_gradients(), to calculate the regression coefficients. It is important to notice that the output of the MultipleLinearRegression function returns the newly created MultipleLinearRegressionModel object type. This is done to encapsulate the trained model parameters into a reusable bundle that contains the fitted coefficients.
We have also implemented a predict() method into the new MultipleLinearRegressionModel struct which should enable us to generate new predictions and forecasts by simply passing the new feature X inputs to the function. This modular approach avoids the need to re-fit the model each time when making new predictions allowing us to store, access, and conveniently manipulate model coefficients.
Once again to ensure that the file is accessible to the compiler we need to also add a reference module. For this, let's create the file named model.cairo under the main src folder and add the following:
// in model.cairomod multiple_linear_regression_model;
Helper functions
Now let's create an additional file named helper_functions.cairo under the main src folder which will host all our helper functions required to construct the MLR Solver. Some of these functions stored here will also be used later during the testing phase to assess the model's performance once fitted. This file consists of multiple functions some of which include:
Function to help with retrieving tensor data by row and column index, which are essential for MLR construction
Function to compute the accuracy of our model using the R-squared method
A function for computing tensor means used in testing.
Functions dedicated to normalizing feature inputs, enabling accurate predictions and forecasts.
A rescaling function tailored to adjust prediction results to appropriate sizes.
use debug::PrintTrait;use array::{ArrayTrait, SpanTrait};use orion::operators::tensor::{Tensor, TensorTrait, FP16x16Tensor, U32Tensor, U32TensorAdd, FP16x16TensorSub, FP16x16TensorAdd,FP16x16TensorDiv, FP16x16TensorMul};use orion::numbers::{FP16x16, FixedTrait};// retrieves row data by index in a 2D tensorfnget_tensor_data_by_row(tensor_data:Tensor<FP16x16>, row_index:u32,) ->Tensor<FP16x16> {let column_len =*tensor_data.shape.at(1); //13// create new arrayletmut result =ArrayTrait::<FP16x16>::new();// loop through the x values and append values letmut i:u32=0;loop {if i >= column_len {break (); } result.append(tensor_data.at(indices:array![row_index, i].span())); i +=1; };let resultant_tensor =TensorTrait::<FP16x16 >::new(array![column_len].span(), data: result.span());return resultant_tensor;}// transposes tensorfntranspose_tensor(tensor_data:Tensor<FP16x16>) ->Tensor<FP16x16> {let tensor_transposed = tensor_data.transpose(axes:array![1, 0].span());return tensor_transposed;}fncalculate_mean(tensor_data:Tensor<FP16x16>) ->FP16x16 {let tensor_size =FixedTrait::<FP16x16>::new_unscaled(tensor_data.data.len(), false);let cumulated_sum = tensor_data.cumsum(0, Option::None(()), Option::None(()));let sum_result = cumulated_sum.data[tensor_data.data.len() -1];let mean =*sum_result / tensor_size;return mean;}// Calculates the R-squared score between two tensors.fncalculate_r_score(Y_values:Tensor<FP16x16>, Y_pred_values:Tensor<FP16x16>) ->FP16x16 {letmutY_values_=Y_values;let mean_y_value =calculate_mean(Y_values);// creating the appropriate tensor shapes and empty arrays to populate values intoletmut squared_diff_shape = array::ArrayTrait::new(); squared_diff_shape.append(Y_values.data.len());letmut squared_diff_vals = array::ArrayTrait::new();letmut squared_mean_diff_shape = array::ArrayTrait::new(); squared_mean_diff_shape.append(Y_values.data.len());letmut squared_mean_diff_vals = array::ArrayTrait::new();letmut i:u32=0;loop {matchY_values_.data.pop_front() {Option::Some(y_value) => {let diff_pred =*y_value -*Y_pred_values.data.at(i);let squared_diff = diff_pred * diff_pred; squared_diff_vals.append(squared_diff);let diff_mean =*y_value - mean_y_value;let squared_mean_diff = diff_mean * diff_mean; squared_mean_diff_vals.append(squared_mean_diff); i +=1; },Option::None=> { break; } } };let squared_diff_tensor =TensorTrait::<FP16x16 >::new(squared_diff_shape.span(), squared_diff_vals.span());let squared_mean_diff_tensor =TensorTrait::<FP16x16 >::new(squared_mean_diff_shape.span(), squared_mean_diff_vals.span());let sum_squared_diff = squared_diff_tensor.cumsum(0, Option::None(()), Option::None(()));let sum_squared_mean_diff = squared_mean_diff_tensor.cumsum(0, Option::None(()), Option::None(()));let r_score =FixedTrait::new_unscaled(1, false)-*sum_squared_diff.data.at(Y_values.data.len() -1)/*sum_squared_mean_diff.data.at(Y_values.data.len() -1);return r_score;}// computes the x_min, x_max and x_range. Used for helping in normalizing and denormalizing user input values operations
fnnormalize_user_x_inputs( x_inputs:Tensor<FP16x16>, original_x_values:Tensor<FP16x16>) ->Tensor<FP16x16> {letmut x_inputs_normalized =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![FixedTrait::new(10, false)].span());letmut x_min =ArrayTrait::<FP16x16>::new();letmut x_max =ArrayTrait::<FP16x16>::new();letmut x_range =ArrayTrait::<FP16x16>::new();letmut result =ArrayTrait::<FP16x16>::new();if original_x_values.shape.len() > 1 {let transposed_tensor = original_x_values.transpose(axes:array![1, 0].span());let data_len =*transposed_tensor.shape.at(0); //13// loop through each row calculating the min, max, and range row values for each feature columnletmut i:u32=0;loop {if i >= data_len {break (); }letmut transposed_tensor_row =get_tensor_data_by_row(transposed_tensor, i); x_min.append(transposed_tensor_row.min_in_tensor()); x_max.append(transposed_tensor_row.max_in_tensor()); x_range.append( transposed_tensor_row.max_in_tensor() - transposed_tensor_row.min_in_tensor() ); i +=1; };letmut x_min_tensor =TensorTrait::new(shape:array![data_len].span(), data: x_min.span());letmut x_max_tensor =TensorTrait::new(shape:array![data_len].span(), data: x_max.span());letmut x_range_tensor =TensorTrait::new( shape:array![data_len].span(), data: x_range.span() );// for normalizing 2D user inputted feature valsif x_inputs.shape.len() > 1 {letmut j:u32=0;loop {if j >=*x_inputs.shape.at(0) {break (); };letmut row_data =get_tensor_data_by_row(x_inputs, j);letmut norm_row_data = (row_data - x_min_tensor) / x_range_tensor;letmut k:u32=0;loop {if k >= norm_row_data.data.len() {break (); }; result.append(*norm_row_data.data.at(k)); k +=1; }; j +=1; }; x_inputs_normalized =TensorTrait::<FP16x16 >::new(array![*x_inputs.shape.at(0), *x_inputs.shape.at(1)].span(), data: result.span() ); };// for normalizing 1D feature inputif x_inputs.shape.len() ==1 { x_inputs_normalized = (x_inputs - x_min_tensor) / x_range_tensor; }; }if original_x_values.shape.len() ==1 {letmut x_min_tensor =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![original_x_values.min_in_tensor()].span());letmut x_max_tensor =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![original_x_values.max_in_tensor()].span());letmut x_range_tensor =TensorTrait::<FP16x16 >::new( shape:array![1].span(), data:array![original_x_values.max_in_tensor() - original_x_values.min_in_tensor()].span() );letmut diff = ((x_inputs - x_min_tensor)); x_inputs_normalized = ((x_inputs - x_min_tensor)) / x_range_tensor; };return x_inputs_normalized;}// rescales model predictions to standard formatfnrescale_predictions( prediction_result:Tensor<FP16x16>, y_values:Tensor<FP16x16>) ->Tensor<FP16x16> {letmut rescale_predictions =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![FixedTrait::new(10, false)].span());letmut y_min_array =ArrayTrait::<FP16x16>::new();letmut y_max_array =ArrayTrait::<FP16x16>::new();letmut y_range_array =ArrayTrait::<FP16x16>::new();letmut y_max = y_values.max_in_tensor();letmut y_min = y_values.min_in_tensor();letmut y_range = y_values.max_in_tensor() - y_values.min_in_tensor();// convert to tensor format for ease of math operationslet y_min_tensor =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![y_min].span());let y_max_tensor =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![y_max].span());let y_range_tensor =TensorTrait::<FP16x16 >::new(shape:array![1].span(), data:array![y_range].span()); rescale_predictions = (prediction_result * y_range_tensor) + y_min_tensor;return rescale_predictions;}
Running tests on the model
At this stage, we have already implemented all the important sections of this tutorial in Cairo. What's left is doing some testing to ensure our model is behaving as expected. To perform our test we will create a new test file called test.cairo under the main src folder and import all the necessary libraries including our x and y values and the MLR solver traits and functions as seen below.
// use traits::Into;use debug::PrintTrait;use array::{ArrayTrait, SpanTrait};use multiple_linear_regresion::datasets::aave_data::aave_x_features::aave_x_features;use multiple_linear_regresion::datasets::aave_data::aave_y_labels::aave_y_labels; use multiple_linear_regresion::datasets::user_inputs_data::aave_weth_revenue_data_input::{aave_weth_revenue_data_input };
use multiple_linear_regresion::model::multiple_linear_regression_model::{MultipleLinearRegressionModel, MultipleLinearRegression, MultipleLinearRegressionModelTrait};use multiple_linear_regresion::data_preprocessing::{Dataset, DatasetTrait};use multiple_linear_regresion::helper_functions::{get_tensor_data_by_row, transpose_tensor, calculate_mean , calculate_r_score, normalize_user_x_inputs, rescale_predictions};use orion::numbers::{FP16x16, FixedTrait};use orion::operators::tensor::{Tensor, TensorTrait, FP16x16Tensor, U32Tensor, U32TensorAdd, FP16x16TensorSub, FP16x16TensorAdd, FP16x16TensorDiv, FP16x16TensorMul};#[test]#[available_gas(99999999999999999)]fnmultiple_linear_regression_test() {//Constructing our modelletmut main_x_vals =aave_x_features();letmut main_y_vals =aave_y_labels();letmut dataset =Dataset{x_values: main_x_vals,y_values:main_y_vals};letmut normalized_dataset = dataset.normalize_dataset();letmut model =MultipleLinearRegression(normalized_dataset);letmut model_coefficients = model.coefficients;letmut reconstructed_ys = model.predict (normalized_dataset.x_values);letmut r_squared_score =calculate_r_score(normalized_dataset.y_values,reconstructed_ys);// r_squared_score.print(); // model accuracy around 0.9969482421875 // checking if data has been normalized correctlyassert(normalized_dataset.x_values.max_in_tensor() <= FixedTrait::new(65536, false), 'normalized x not between 0-1');assert(normalized_dataset.x_values.min_in_tensor() >= FixedTrait::new(0, false), 'normalized x not between 0-1');assert(normalized_dataset.y_values.max_in_tensor() <= FixedTrait::new(65536, false), 'normalized y not between 0-1');assert(normalized_dataset.x_values.min_in_tensor() >= FixedTrait::new(0, false), 'normalized y not between 0-1');// performing checks on the shape of normalized dataassert(normalized_dataset.x_values.data.len()== main_x_vals.data.len() && normalized_dataset.y_values.data.len()== main_y_vals.data.len() , 'normalized data shape mismatch');// performing checks on the shape of coefficient values (gradient vals + bias)assert(model.coefficients.data.len() == *main_x_vals.shape.at(1)+1, 'coefficient data shape mismatch');// model accuracy deviance checksassert(r_squared_score >= FixedTrait::new(62259, false), 'AAVE model acc. less than 95%');// using the model to forecast aave's 7-day WETH net lifetime repayments forecast let last_7_days_aave_data =aave_weth_revenue_data_input();let last_7_days_aave_data_normalized =normalize_user_x_inputs(last_7_days_aave_data, main_x_vals );letmut forecast_results = model.predict (last_7_days_aave_data_normalized); let mut rescale_forecasts = rescale_predictions(forecast_results, main_y_vals); // PS. ** the rescaled forecasted outputs are denominated in thousands of ETH
// (*rescale_forecasts.data.at(0)).print(); // day1 forecast: 95.66773986816406 // (*rescale_forecasts.data.at(1)).print(); // day2: 96.64869689941406// (*rescale_forecasts.data.at(5)).print(); // day6: 99.44300842285156// (*rescale_forecasts.data.at(6)).print(); // day7: 100.57145690917969}
Our model will get tested under the multiple_linear_regression_test() function which will follow these steps:
Data retrieval: The function initiates by fetching the AAVE dataset's x and y values.
Dataset construction and normalization: A new Dataset object gets initialized by passing the x and y variables. It is then normalized using the built-in normalize_dataset() method.
Model fitting: Using the MultipleLinearRegression function we fit the normalized dataset and compute the regression coefficients.
Computing accuracy of the model: To calculate the accuracy we utilize the predict method to compute the dot product between the model's regression coefficients and the x values. We then compute the R-squared score to measure the accuracy of our model.
Perform some testing: In the subsequent step we perform some checks to ensure that the tensor shape/dimension is correct. We also check the model's accuracy deviance to see if it's still within an acceptable range.
Making forecasts: If our checks have passed then our model should be clear to enable us to make new predictions. For this, we will use the aave_weth_revenue_data_input() values which represent the most recent AAVE datapoints which should enable us to make forecasts for the upcoming 7 days of AAVE's WETH Pool Lifetime Repayments.
🥁.... And as we can our test cases have passed! Hooray!! 🎊
Congratulations on reaching this point! 👏 You are now ready to implement fully transparent and verifiable forecasting solutions using this MLR framework.
If you're looking for more examples of using the MLR Solver, look into here as it covers more easy-to-follow jupyter notebook tutorials (e.g. Boston dataset). 😁
We invite you to join us in forging a future by making AI a transparent and reliable resource for all!
