MedDeepBit.h

#pragma once
#include <MedAlgo/MedAlgo/MedAlgo.h>
#include <algorithm>
#define LOCAL_SECTION LOG_MEDALGO
#define LOCAL_LEVEL LOG_DEF_LEVEL
extern MedLogger global_logger;
 
//Deep Bit
 
struct MedDeepBitParams : public SerializableObject {
    int max_depth;
    int num_iterations;
    int num_ftrs_per_round;
    int num_vals_to_be_categorial;
    int nparts_auc;
    int niter_auc_gitter;
    int niter_coordinate_descent;
    int internal_test_ratio;
    double fraction_auc;
    double grid_fraction;
    double min_fraction_zeros_ones;
    double frac_continuous_frequent;
    double frac_categorial_frequent;
    double lambda;
    double min_cor_bin_ftr;
 
    MedDeepBitParams() {
        max_depth = 20; num_iterations = 100; num_ftrs_per_round = 1000; num_vals_to_be_categorial = 10; nparts_auc = 1; niter_coordinate_descent = 0;
        niter_auc_gitter = 0;  grid_fraction = 0.003; min_fraction_zeros_ones = 0.003;
        frac_continuous_frequent = 0.1; frac_categorial_frequent = 0.1; lambda = 20; fraction_auc = 1; internal_test_ratio = 10, min_cor_bin_ftr = 0.05;
    }
    ADD_CLASS_NAME(MedDeepBitParams)
    size_t get_size();
    size_t serialize(unsigned char *blob);
    size_t deserialize(unsigned char *blob);
    string to_string();
};
 
MEDSERIALIZE_SUPPORT(MedDeepBitParams)
 
class MedDeepBit : public MedPredictor {    
 
public:
    MedDeepBitParams params;
    void init_defaults();
    int init(void *classifier_params) { this->params = *((MedDeepBitParams*)classifier_params); return 0; };
    int set_params(map<string, string>& initialization_map);
    MedDeepBit() { classifier_type = MODEL_DEEP_BIT; init_defaults(); }
    ~MedDeepBit() {}
    MedDeepBit(void *_params) { params = *(MedDeepBitParams *)_params; }
    MedDeepBit(MedDeepBitParams& _params) { params = _params; }
    int Learn(float *x, float *y, const float *w, int nsamples, int nftrs);
    int Learn(float *x, float *y, int nsamples, int nftrs);
    int Predict(float *x, float *&preds, int nsamples, int nftrs) const;
    virtual void print(FILE *fp, const string& prefix, int level=0) const;
    ADD_CLASS_NAME(MedDeepBit)
    size_t get_size();
    size_t serialize(unsigned char *blob);
    size_t deserialize(unsigned char *blob);
    void print_model(FILE *fp, const string& prefix) const;
 
private:
    vector<vector<double>> x, internal_test_x, internal_test_transposed_x;
    vector<double> y, scores, r, internal_test_scores, avx, sdx;
    vector<int> label, internal_test_label;
    int nftrs, nsamples, internal_test_nsamples, num_bin_ftrs;
    vector<string> ftr_names;
    double avy;
 
    vector<vector<double>> ftr_grids;
    vector<int> is_categorial;
    vector<vector<double>> frequent_ftr_vals;
    // For each binary feature we document the raw feature number, a spceific value val of that feature and a boolean value bool_val. 
    // For categorial featues, we ask weather a given sample equals val or not, for bool_val in {true, false}, correpondingly.
    // For quantitative features, we ask weather a given sample is above or under val for the values of true and false for bool_val, correspondingly.
    vector<tuple<int, int, bool, bool>> bin_ftrs_map;
    vector<vector<char>> bin_ftrs;
    vector<vector<int>> bin_ftr_indexes;
    vector<vector<double>> bin_ftr_avg_sd_beta;
 
    void print_log();
    void train_model();
    void predict_train() {}
    double predict_sample(const vector<double>& x, int niter) const;
    double predict_sample(const vector<double>& x) const;
    void get_normalized_col(const vector<char>& col, vector<double>& normalized_col, double& av, double& std);
    double perform_lasso_iteration(const vector<double>& xk_train, const vector<double>& r, double lambda, double alpha);
    void get_col_without_na(const vector<double>& col, vector<double>& col_without_na);
    void get_avgs(const vector<vector<double>>& x, vector<double>& avx);
    void get_sds(const vector<vector<double>>& x, vector<double>& sdx);
    double avg(const vector<char>& vec);
    double avg(const vector<double>& vec);
    double sd(const vector<char>& binary_vec);
    double sd(const vector<double>& vec);
    void transpose(const vector<vector<double>>& before, vector<vector<double>>& after);
    void init(float *x1, float *y1, int nsamples1, int nftrs1);
    void predict(const vector<vector<double>>& x, vector<double>& scores) const;
    double get_normalized_val(double x_val, int j);
    void impute_x(vector<vector<double>>& x, const vector<double>& avx);
    bool is_viable_01_ratios(int count0, int count1, int count_pos0, int count_pos1);
    bool is_bin_ftr_valid(const vector<char>& bin_ftr);
    void make_bin_ftrs(int j, const vector<double>& vals, bool is_categorial);
    int step_function(int step, int i);
    void mark_grids_and_frequent_vals();
    void get_categorial_bin_ftr(const vector<double>& col, double val, bool direction, vector<char>& bin_ftr);
    void get_quant_bin_ftr(const vector<double>& col, double val, bool direction, vector<char>& bin_ftr);
    int get_categorial_bit(double x_val, double val, bool direction) const;
    int get_quant_bit(double x_val, double val, bool direction) const;
    double get_ftr_score(const vector<char>& bin_ftr);
    void mult_bin_ftrs(const vector<char>& ftr1, vector<char>& ftr2);
    void get_bin_ftr(int bin_ftr_index, vector<char>& bin_ftr);
    void print_ftr_characteristics(int index);
    void calc_bin_ftr_scores(const vector<char>& bin_ftr, double& av, double& std, double& b, vector<double>& scores1, vector<double>& r1, bool is_full_step = false);
    void gen_random_indexes(vector<int>& random_indexes);
    void score_random_ftrs(vector<double>& ftr_scores, const vector<int>& random_indexes, const vector<char>& final_bin_ftr);
    void get_bin_ftr_of_it(int it, vector<char>& bin_ftr);
    void do_auc_gittering();
    void do_coordinate_descent(int num_iterations_descent);
    double special_auc(const vector<double>& all_predictions, const vector<int>& all_label, bool is_weighted = false, int nparts = 1);
};
 
MEDSERIALIZE_SUPPORT(MedDeepBit)