import streamlit as st import pandas as pd import plotly.express as px import plotly.graph_objects as go from Eda_functions import format_numbers import numpy as np import pickle from st_aggrid import AgGrid from st_aggrid import GridOptionsBuilder,GridUpdateMode from utilities import set_header,load_local_css from st_aggrid import GridOptionsBuilder import time import itertools import statsmodels.api as sm import numpy as npc import re import itertools from sklearn.metrics import mean_absolute_error, r2_score,mean_absolute_percentage_error from sklearn.preprocessing import MinMaxScaler import os import matplotlib.pyplot as plt from statsmodels.stats.outliers_influence import variance_inflation_factor st.set_option('deprecation.showPyplotGlobalUse', False) import statsmodels.api as sm import statsmodels.formula.api as smf from datetime import datetime import seaborn as sns from Data_prep_functions import * def get_random_effects(media_data, panel_col, mdf): random_eff_df = pd.DataFrame(columns=[panel_col, "random_effect"]) for i, market in enumerate(media_data[panel_col].unique()): print(i, end='\r') intercept = mdf.random_effects[market].values[0] random_eff_df.loc[i, 'random_effect'] = intercept random_eff_df.loc[i, panel_col] = market return random_eff_df def mdf_predict(X, mdf, random_eff_df) : X['fixed_effect'] = mdf.predict(X) merged_df=pd.merge(X[[panel_col,target_col]], random_eff_df, on = panel_col, how = 'left') X['random_effect'] = merged_df['random_effect'] X['pred'] = X['fixed_effect'] + X['random_effect'] return X['pred'] st.set_page_config( page_title="Model Build", page_icon=":shark:", layout="wide", initial_sidebar_state='collapsed' ) load_local_css('styles.css') set_header() st.title('1. Build Your Model') panel_col= 'markets' # set the panel column date_col = 'date' target_col = 'total_approved_accounts_revenue' media_data=pd.read_csv('upf_data_converted.csv') media_data.columns=[i.lower().replace('-','').replace(':','').replace("__", "_") for i in media_data.columns] # st.write(media_data.columns) media_data.sort_values(date_col, inplace=True) media_data.reset_index(drop=True,inplace=True) date=media_data[date_col] st.session_state['date']=date revenue=media_data[target_col] media_data.drop([target_col],axis=1,inplace=True) media_data.drop([date_col],axis=1,inplace=True) media_data.reset_index(drop=True,inplace=True) if st.toggle('Apply Transformations on DMA/Panel Level'): dma=st.selectbox('Select the Level of data ',[ col for col in media_data.columns if col.lower() in ['dma','panel', 'markets']]) else: #""" code to aggregate data on date """ dma=None # dma_dict={ dm:media_data[media_data[dma]==dm] for dm in media_data[dma].unique()} # st.write(dma_dict) st.markdown('## Select the Range of Transformations') columns = st.columns(2) old_shape=media_data.shape if "old_shape" not in st.session_state: st.session_state['old_shape']=old_shape with columns[0]: slider_value_adstock = st.slider('Select Adstock Range (only applied to media)', 0.0, 1.0, (0.2, 0.4), step=0.1, format="%.2f") with columns[1]: slider_value_lag = st.slider('Select Lag Range (applied to media, seasonal, macroeconomic variables)', 1, 7, (1, 3), step=1) # with columns[2]: # slider_value_power=st.slider('Select Power range (only applied to media )',0,4,(1,2),step=1) # with columns[1]: # st.number_input('Select the range of half saturation point ',min_value=1,max_value=5) # st.number_input('Select the range of ') # Section 1 - Transformations Functions def lag(data,features,lags,dma=None): if dma: transformed_data=pd.concat([data.groupby([dma])[features].shift(lag).add_suffix(f'_lag_{lag}') for lag in lags],axis=1) transformed_data=transformed_data.fillna(method='bfill') return pd.concat([transformed_data,data],axis=1) else: #''' data should be aggregated on date''' transformed_data=pd.concat([data[features].shift(lag).add_suffix(f'_lag_{lag}') for lag in lags],axis=1) transformed_data=transformed_data.fillna(method='bfill') return pd.concat([transformed_data,data],axis=1) #adstock def adstock(df, alphas, cutoff, features,dma=None): # st.write(features) if dma: transformed_data=pd.DataFrame() for d in df[dma].unique(): dma_sub_df = df[df[dma] == d] n = len(dma_sub_df) weights = np.array([[[alpha**(i-j) if i >= j and j >= i-cutoff else 0. for j in range(n)] for i in range(n)] for alpha in alphas]) X = dma_sub_df[features].to_numpy() res = pd.DataFrame(np.hstack(weights @ X), columns=[f'{col}_adstock_{alpha}' for alpha in alphas for col in features]) transformed_data=pd.concat([transformed_data,res],axis=0) transformed_data.reset_index(drop=True,inplace=True) return pd.concat([transformed_data,df],axis=1) else: n = len(df) weights = np.array([[[alpha**(i-j) if i >= j and j >= i-cutoff else 0. for j in range(n)] for i in range(n)] for alpha in alphas]) X = df[features].to_numpy() res = pd.DataFrame(np.hstack(weights @ X), columns=[f'{col}_adstock_{alpha}' for alpha in alphas for col in features]) return pd.concat([res,df],axis=1) # Section 2 - Begin Transformations if 'media_data' not in st.session_state: st.session_state['media_data']=pd.DataFrame() # variables_to_be_transformed=[col for col in media_data.columns if col.lower() not in ['dma','panel'] ] # change for buckets variables_to_be_transformed=[col for col in media_data.columns if '_clicks' in col.lower() or '_impress' in col.lower()] # srishti - change # st.write(variables_to_be_transformed) # st.write(media_data[variables_to_be_transformed].dtypes) with columns[0]: if st.button('Apply Transformations'): with st.spinner('Applying Transformations'): transformed_data_lag=lag(media_data,features=variables_to_be_transformed,lags=np.arange(slider_value_lag[0],slider_value_lag[1]+1,1),dma=dma) # variables_to_be_transformed=[col for col in list(transformed_data_lag.columns) if col not in ['Date','DMA','Panel']] #change for buckets variables_to_be_transformed = [col for col in media_data.columns if '_clicks' in col.lower() or '_impress' in col.lower()] # srishti - change transformed_data_adstock=adstock(df=transformed_data_lag, alphas=np.arange(slider_value_adstock[0],slider_value_adstock[1],0.1), cutoff=8, features=variables_to_be_transformed,dma=dma) # st.success('Done') st.success("Transformations complete!") st.write(f'old shape {old_shape}, new shape {transformed_data_adstock.shape}') # st.write(media_data.head(10)) # st.write(transformed_data_adstock.head(10)) transformed_data_adstock.columns = [c.replace(".","_") for c in transformed_data_adstock.columns] # srishti # st.write(transformed_data_adstock.columns) st.session_state['media_data']=transformed_data_adstock # srishti # with st.spinner('Applying Transformations'): # time.sleep(2) # st.success("Transformations complete!") # if st.session_state['media_data'].shape[1]>old_shape[1]: # with columns[0]: # st.write(f'Total no.of variables before transformation: {old_shape[1]}, Total no.of variables after transformation: {st.session_state["media_data"].shape[1]}') #st.write(f'Total no.of variables after transformation: {st.session_state["media_data"].shape[1]}') # Section 3 - Create combinations # bucket=['paid_search', 'kwai','indicacao','infleux', 'influencer','FB: Level Achieved - Tier 1 Impressions', # ' FB: Level Achieved - Tier 2 Impressions','paid_social_others', # ' GA App: Will And Cid Pequena Baixo Risco Clicks', # 'digital_tactic_others',"programmatic" # ] # srishti - bucket names changed bucket=['paid_search', 'kwai','indicacao','infleux', 'influencer','fb_level_achieved_tier_2', 'fb_level_achieved_tier_1','paid_social_others', 'ga_app', 'digital_tactic_others',"programmatic" ] with columns[1]: if st.button('Create Combinations of Variables'): top_3_correlated_features=[] # for col in st.session_state['media_data'].columns[:19]: original_cols = [c for c in st.session_state['media_data'].columns if "_clicks" in c.lower() or "_impressions" in c.lower()] original_cols = [c for c in original_cols if "_lag" not in c.lower() and "_adstock" not in c.lower()] # st.write(original_cols) # for col in st.session_state['media_data'].columns[:19]: for col in original_cols: # srishti - new corr_df=pd.concat([st.session_state['media_data'].filter(regex=col), revenue],axis=1).corr()[target_col].iloc[:-1] top_3_correlated_features.append(list(corr_df.sort_values(ascending=False).head(2).index)) # st.write(col, top_3_correlated_features) flattened_list = [item for sublist in top_3_correlated_features for item in sublist] # all_features_set={var:[col for col in flattened_list if var in col] for var in bucket} all_features_set={var:[col for col in flattened_list if var in col] for var in bucket if len([col for col in flattened_list if var in col])>0} # srishti channels_all=[values for values in all_features_set.values()] # st.write(channels_all) st.session_state['combinations'] = list(itertools.product(*channels_all)) # if 'combinations' not in st.session_state: # st.session_state['combinations']=combinations_all st.session_state['final_selection']=st.session_state['combinations'] st.success('Done') # st.write(f"{len(st.session_state['combinations'])} combinations created") revenue.reset_index(drop=True,inplace=True) if 'Model_results' not in st.session_state: st.session_state['Model_results']={'Model_object':[], 'Model_iteration':[], 'Feature_set':[], 'MAPE':[], 'R2':[], 'ADJR2':[] } def reset_model_result_dct(): st.session_state['Model_results']={'Model_object':[], 'Model_iteration':[], 'Feature_set':[], 'MAPE':[], 'R2':[], 'ADJR2':[] } # if st.button('Build Model'): if 'iterations' not in st.session_state: st.session_state['iterations']=0 # st.write("1",st.session_state["final_selection"]) if 'final_selection' not in st.session_state: st.session_state['final_selection']=False save_path = r"Model/" with columns[1]: if st.session_state['final_selection']: st.write(f'Total combinations created {format_numbers(len(st.session_state["final_selection"]))}') if st.checkbox('Build all iterations'): iterations=len(st.session_state['final_selection']) else: iterations = st.number_input('Select the number of iterations to perform', min_value=0, step=10, value=st.session_state['iterations'],on_change=reset_model_result_dct) # st.write("iterations=", iterations) if st.button('Build Model',on_click=reset_model_result_dct): st.session_state['iterations']=iterations # st.write("2",st.session_state["final_selection"]) # Section 4 - Model st.session_state['media_data']=st.session_state['media_data'].fillna(method='ffill') st.markdown( 'Data Split -- Training Period: May 9th, 2023 - October 5th,2023 , Testing Period: October 6th, 2023 - November 7th, 2023 ') progress_bar = st.progress(0) # Initialize the progress bar # time_remaining_text = st.empty() # Create an empty space for time remaining text start_time = time.time() # Record the start time progress_text = st.empty() # time_elapsed_text = st.empty() # for i, selected_features in enumerate(st.session_state["final_selection"][40000:40000 + int(iterations)]): # st.write(st.session_state["final_selection"]) # for i, selected_features in enumerate(st.session_state["final_selection"]): for i, selected_features in enumerate(st.session_state["final_selection"][0:int(iterations)]): # srishti print("@@@@@@@@@@@@@",i) df = st.session_state['media_data'] fet = [var for var in selected_features if len(var) > 0] inp_vars_str = " + ".join(fet) # new X = df[fet] y = revenue ss = MinMaxScaler() X = pd.DataFrame(ss.fit_transform(X), columns=X.columns) # X = sm.add_constant(X) X['total_approved_accounts_revenue'] = revenue # new X[panel_col] = df[panel_col] X_train = X.iloc[:8000] X_test = X.iloc[8000:] y_train = y.iloc[:8000] y_test = y.iloc[8000:] print(X_train.shape) # model = sm.OLS(y_train, X_train).fit() md = smf.mixedlm("total_approved_accounts_revenue ~ {}".format(inp_vars_str), data=X_train[['total_approved_accounts_revenue'] + fet], groups=X_train[panel_col]) mdf = md.fit() predicted_values = mdf.fittedvalues # st.write(fet) # positive_coeff=fet # negetive_coeff=[] coefficients = mdf.fe_params.to_dict() model_possitive = [col for col in coefficients.keys() if coefficients[col] > 0] # st.write(positive_coeff) # st.write(model_possitive) pvalues = [var for var in list(mdf.pvalues) if var <= 0.06] # if (len(model_possitive) / len(selected_features)) > 0.9 and (len(pvalues) / len(selected_features)) >= 0.8: if (len(model_possitive) / len(selected_features)) > 0 and (len(pvalues) / len(selected_features)) >= 0: # srishti - changed just for testing, revert later # predicted_values = model.predict(X_train) mape = mean_absolute_percentage_error(y_train, predicted_values) r2 = r2_score(y_train, predicted_values) adjr2 = 1 - (1 - r2) * (len(y_train) - 1) / (len(y_train) - len(selected_features) - 1) filename = os.path.join(save_path, f"model_{i}.pkl") with open(filename, "wb") as f: pickle.dump(mdf, f) # with open(r"C:\Users\ManojP\Documents\MMM\simopt\Model\model.pkl", 'rb') as file: # model = pickle.load(file) st.session_state['Model_results']['Model_object'].append(filename) st.session_state['Model_results']['Model_iteration'].append(i) st.session_state['Model_results']['Feature_set'].append(fet) st.session_state['Model_results']['MAPE'].append(mape) st.session_state['Model_results']['R2'].append(r2) st.session_state['Model_results']['ADJR2'].append(adjr2) current_time = time.time() time_taken = current_time - start_time time_elapsed_minutes = time_taken / 60 completed_iterations_text = f"{i + 1}/{iterations}" progress_bar.progress((i + 1) / int(iterations)) progress_text.text(f'Completed iterations: {completed_iterations_text},Time Elapsed (min): {time_elapsed_minutes:.2f}') st.write(f'Out of {st.session_state["iterations"]} iterations : {len(st.session_state["Model_results"]["Model_object"])} valid models') pd.DataFrame(st.session_state['Model_results']).to_csv('model_output.csv') def to_percentage(value): return f'{value * 100:.1f}%' st.title('2. Select Models') if 'tick' not in st.session_state: st.session_state['tick']=False if st.checkbox('Show results of top 10 models (based on MAPE and Adj. R2)',value=st.session_state['tick']): st.session_state['tick']=True st.write('Select one model iteration to generate performance metrics for it:') data=pd.DataFrame(st.session_state['Model_results']) data.sort_values(by=['MAPE'],ascending=False,inplace=True) data.drop_duplicates(subset='Model_iteration',inplace=True) top_10=data.head(10) top_10['Rank']=np.arange(1,len(top_10)+1,1) top_10[['MAPE','R2','ADJR2']]=np.round(top_10[['MAPE','R2','ADJR2']],4).applymap(to_percentage) top_10_table = top_10[['Rank','Model_iteration','MAPE','ADJR2','R2']] #top_10_table.columns=[['Rank','Model Iteration Index','MAPE','Adjusted R2','R2']] gd=GridOptionsBuilder.from_dataframe(top_10_table) gd.configure_pagination(enabled=True) gd.configure_selection(use_checkbox=True) gridoptions=gd.build() table = AgGrid(top_10,gridOptions=gridoptions,update_mode=GridUpdateMode.SELECTION_CHANGED) selected_rows=table.selected_rows # if st.session_state["selected_rows"] != selected_rows: # st.session_state["build_rc_cb"] = False st.session_state["selected_rows"] = selected_rows if 'Model' not in st.session_state: st.session_state['Model']={} if len(selected_rows)>0: st.header('2.1 Results Summary') model_object=data[data['Model_iteration']==selected_rows[0]['Model_iteration']]['Model_object'] features_set=data[data['Model_iteration']==selected_rows[0]['Model_iteration']]['Feature_set'] with open(str(model_object.values[0]), 'rb') as file: # print(file) model = pickle.load(file) st.write(model.summary()) st.header('2.2 Actual vs. Predicted Plot') df=st.session_state['media_data'] X=df[features_set.values[0]] # X = sm.add_constant(X) y=revenue ss = MinMaxScaler() X = pd.DataFrame(ss.fit_transform(X), columns=X.columns) X['total_approved_accounts_revenue'] = revenue # new X[panel_col] = df[panel_col] X_train=X.iloc[:8000] X_test=X.iloc[8000:] y_train=y.iloc[:8000] y_test=y.iloc[8000:] st.session_state['X']=X_train st.session_state['features_set']=features_set.values[0] metrics_table,line,actual_vs_predicted_plot=plot_actual_vs_predicted(date, y_train, model.fittedvalues, model,target_column='Revenue') st.plotly_chart(actual_vs_predicted_plot,use_container_width=True) random_eff_df = get_random_effects(media_data, panel_col, model) st.markdown('## 2.3 Residual Analysis') columns=st.columns(2) with columns[0]: fig=plot_residual_predicted(y_train,model.fittedvalues,X_train) st.plotly_chart(fig) with columns[1]: st.empty() fig = qqplot(y_train,model.fittedvalues) st.plotly_chart(fig) with columns[0]: fig=residual_distribution(y_train,model.fittedvalues) st.pyplot(fig) vif_data = pd.DataFrame() # X=X.drop('const',axis=1) vif_data["Variable"] = X_train.columns vif_data["VIF"] = [variance_inflation_factor(X_train.values, i) for i in range(X_train.shape[1])] vif_data.sort_values(by=['VIF'],ascending=False,inplace=True) vif_data=np.round(vif_data) vif_data['VIF']=vif_data['VIF'].astype(float) st.header('2.4 Variance Inflation Factor (VIF)') #st.dataframe(vif_data) color_mapping = { 'darkgreen': (vif_data['VIF'] < 3), 'orange': (vif_data['VIF'] >= 3) & (vif_data['VIF'] <= 10), 'darkred': (vif_data['VIF'] > 10) } # Create a horizontal bar plot fig, ax = plt.subplots() fig.set_figwidth(10) # Adjust the width of the figure as needed # Sort the bars by descending VIF values vif_data = vif_data.sort_values(by='VIF', ascending=False) # Iterate through the color mapping and plot bars with corresponding colors for color, condition in color_mapping.items(): subset = vif_data[condition] bars = ax.barh(subset["Variable"], subset["VIF"], color=color, label=color) # Add text annotations on top of the bars for bar in bars: width = bar.get_width() ax.annotate(f'{width:}', xy=(width, bar.get_y() + bar.get_height() / 2), xytext=(5, 0), textcoords='offset points', va='center') # Customize the plot ax.set_xlabel('VIF Values') #ax.set_title('2.4 Variance Inflation Factor (VIF)') #ax.legend(loc='upper right') # Display the plot in Streamlit st.pyplot(fig) with st.expander('Results Summary Test data'): ss = MinMaxScaler() X_test = pd.DataFrame(ss.fit_transform(X_test), columns=X_test.columns) st.header('2.2 Actual vs. Predicted Plot') metrics_table,line,actual_vs_predicted_plot=plot_actual_vs_predicted(date, y_test, mdf_predict(X_test,mdf, random_eff_df), model,target_column='Revenue') st.plotly_chart(actual_vs_predicted_plot,use_container_width=True) st.markdown('## 2.3 Residual Analysis') columns=st.columns(2) with columns[0]: fig=plot_residual_predicted(revenue,mdf_predict(X_test,mdf, random_eff_df),X_test) st.plotly_chart(fig) with columns[1]: st.empty() fig = qqplot(revenue,mdf_predict(X_test,mdf, random_eff_df)) st.plotly_chart(fig) with columns[0]: fig=residual_distribution(revenue,mdf_predict(X_test,mdf, random_eff_df)) st.pyplot(fig) value=False if st.checkbox('Save this model to tune',key='build_rc_cb'): mod_name=st.text_input('Enter model name') if len(mod_name)>0: st.session_state['Model'][mod_name]={"Model_object":model,'feature_set':st.session_state['features_set'],'X_train':X_train} st.session_state['X_train']=X_train st.session_state['X_test']=X_test st.session_state['y_train']=y_train st.session_state['y_test']=y_test with open("best_models.pkl", "wb") as f: pickle.dump(st.session_state['Model'], f) st.success('Model saved!, Proceed next page to tune the model') value=False