void recommender_base::complete_row(const common::sfv_t& query,
common::sfv_t& ret) const {
ret.clear();
vector<pair<string, float> > ids;
similar_row(query, ids, complete_row_similar_num_);
if (ids.size() == 0) {
return;
}
size_t exist_row_num = 0;
for (size_t i = 0; i < ids.size(); ++i) {
common::sfv_t row;
orig_.get_row(ids[i].first, row);
if (row.size() == 0) {
continue;
} else {
++exist_row_num;
}
float ratio = ids[i].second;
for (size_t j = 0; j < row.size(); ++j) {
ret.push_back(make_pair(row[j].first, row[j].second * ratio));
}
}
if (exist_row_num == 0) {
return;
}
common::sort_and_merge(ret);
for (size_t i = 0; i < ret.size(); ++i) {
ret[i].second /= exist_row_num;
}
}
eigen_svec_t eigen_feature_mapper::convertc(const common::sfv_t& src) const {
eigen_svec_t ret(d_);
for (common::sfv_t::const_iterator it = src.begin(); it != src.end(); ++it) {
insertc(*it, ret);
}
return ret;
}
void confidence_weighted::update(
const common::sfv_t& sfv,
float step_width,
const string& pos_label,
const string& neg_label) {
util::concurrent::scoped_wlock lk(storage_->get_lock());
for (common::sfv_t::const_iterator it = sfv.begin(); it != sfv.end(); ++it) {
const string& feature = it->first;
float val = it->second;
storage::feature_val2_t val2;
storage_->get2_nolock(feature, val2);
storage::val2_t pos_val(0.f, 1.f);
storage::val2_t neg_val(0.f, 1.f);
ClassifierUtil::get_two(val2, pos_label, neg_label, pos_val, neg_val);
const float C = config_.regularization_weight;
float covar_pos_step = 2.f * step_width * val * val * C;
float covar_neg_step = 2.f * step_width * val * val * C;
storage_->set2_nolock(
feature,
pos_label,
storage::val2_t(pos_val.v1 + step_width * pos_val.v2 * val,
1.f / (1.f / pos_val.v2 + covar_pos_step)));
if (neg_label != "") {
storage_->set2_nolock(
feature,
neg_label,
storage::val2_t(neg_val.v1 - step_width * neg_val.v2 * val,
1.f / (1.f / neg_val.v2 + covar_neg_step)));
}
}
touch(pos_label);
}
common::sfv_t scalar_dot(const common::sfv_t& p, double s) {
common::sfv_t ret;
for (common::sfv_t::const_iterator it = p.begin(); it != p.end(); ++it) {
ret.push_back(make_pair((*it).first, (*it).second*s));
}
return ret;
}
double sum2(const common::sfv_t& p) {
double s = 0;
for (common::sfv_t::const_iterator it = p.begin(); it != p.end(); ++it) {
s += std::pow((*it).second, 2);
}
return s;
}
void local_storage::inp(const common::sfv_t& sfv, map_feature_val1_t& ret)
const {
ret.clear();
std::vector<float> ret_id(class2id_.size());
for (common::sfv_t::const_iterator it = sfv.begin(); it != sfv.end(); ++it) {
const string& feature = it->first;
const float val = it->second;
id_features3_t::const_iterator it2 = tbl_.find(feature);
if (it2 == tbl_.end()) {
continue;
}
const id_feature_val3_t& m = it2->second;
for (id_feature_val3_t::const_iterator it3 = m.begin(); it3 != m.end();
++it3) {
ret_id[it3->first] += it3->second.v1 * val;
}
}
for (size_t i = 0; i < ret_id.size(); ++i) {
if (ret_id[i] == 0.f) {
continue;
}
ret[class2id_.get_key(i)] = ret_id[i];
}
}
void arow::update(
const common::sfv_t& sfv,
float alpha,
float beta,
const std::string& pos_label,
const std::string& neg_label) {
storage::storage_base* sto = get_storage();
for (common::sfv_t::const_iterator it = sfv.begin(); it != sfv.end(); ++it) {
const string& feature = it->first;
float val = it->second;
storage::feature_val2_t ret;
sto->get2(feature, ret);
storage::val2_t pos_val(0.f, 1.f);
storage::val2_t neg_val(0.f, 1.f);
ClassifierUtil::get_two(ret, pos_label, neg_label, pos_val, neg_val);
sto->set2(
feature,
pos_label,
storage::val2_t(
pos_val.v1 + alpha * pos_val.v2 * val,
pos_val.v2 - beta * pos_val.v2 * pos_val.v2 * val * val));
if (neg_label != "") {
sto->set2(
feature,
neg_label,
storage::val2_t(
neg_val.v1 - alpha * neg_val.v2 * val,
neg_val.v2 - beta * neg_val.v2 * neg_val.v2 * val * val));
}
}
}
void local_storage::inp(const common::sfv_t& sfv, map_feature_val1_t& ret)
const {
ret.clear();
scoped_rlock lk(mutex_);
// Use uin64_t map instead of string map as hash function for string is slow
jubatus::util::data::unordered_map<uint64_t, double> ret_id;
for (common::sfv_t::const_iterator it = sfv.begin(); it != sfv.end(); ++it) {
const string& feature = it->first;
const double val = it->second;
id_features3_t::const_iterator it2 = tbl_.find(feature);
if (it2 == tbl_.end()) {
continue;
}
const id_feature_val3_t& m = it2->second;
for (id_feature_val3_t::const_iterator it3 = m.begin(); it3 != m.end();
++it3) {
ret_id[it3->first] += it3->second.v1 * val;
}
}
std::vector<std::string> labels = class2id_.get_all_id2key();
for (size_t i = 0; i < labels.size(); ++i) {
const std::string& label = labels[i];
uint64_t id = class2id_.get_id_const(label);
if (id == common::key_manager::NOTFOUND || ret_id.count(id) == 0) {
ret[label] = 0.0;
} else {
ret[label] = ret_id[id];
}
}
}
void storage_base::inp(const common::sfv_t& sfv, map_feature_val1_t& ret)
const {
ret.clear();
for (common::sfv_t::const_iterator it = sfv.begin(); it != sfv.end(); ++it) {
const string& feature = it->first;
const float val = it->second;
feature_val1_t fval1;
get(feature, fval1);
for (feature_val1_t::const_iterator it2 = fval1.begin(); it2 != fval1.end();
++it2) {
ret[it2->first] += it2->second * val;
}
}
}
void recommender_base::complete_row(const std::string& id,
common::sfv_t& ret) const {
ret.clear();
common::sfv_t sfv;
orig_.get_row(id, sfv);
complete_row(sfv, ret);
}
float recommender_base::calc_l2norm(const common::sfv_t& query) {
float ret = 0.f;
for (size_t i = 0; i < query.size(); ++i) {
ret += query[i].second * query[i].second;
}
return sqrt(ret);
}
float inverted_index_storage::calc_l2norm(const common::sfv_t& sfv) {
float ret = 0.f;
for (size_t i = 0; i < sfv.size(); ++i) {
ret += sfv[i].second * sfv[i].second;
}
return std::sqrt(ret);
}
void inverted_index_storage::calc_scores(
const common::sfv_t& query,
vector<pair<string, float> >& scores,
size_t ret_num) const {
float query_norm = calc_l2norm(query);
if (query_norm == 0.f) {
return;
}
jubatus::util::data::unordered_map<uint64_t, float> i_scores;
for (size_t i = 0; i < query.size(); ++i) {
const string& fid = query[i].first;
float val = query[i].second;
add_inp_scores(fid, val, i_scores);
}
vector<pair<float, uint64_t> > sorted_scores;
for (jubatus::util::data::unordered_map<uint64_t, float>::
const_iterator it = i_scores.begin(); it != i_scores.end(); ++it) {
float norm = calc_columnl2norm(it->first);
float normed_score = (norm != 0.f) ? it->second / norm / query_norm : 0.f;
sorted_scores.push_back(make_pair(normed_score, it->first));
}
sort(sorted_scores.rbegin(), sorted_scores.rend());
for (size_t i = 0; i < sorted_scores.size() && i < ret_num; ++i) {
scores.push_back(
make_pair(column2id_.get_key(sorted_scores[i].second),
sorted_scores[i].first));
}
}
static float squared_norm(const common::sfv_t& fv) {
float norm = 0.f;
for (size_t i = 0; i < fv.size(); ++i) {
norm += fv[i].second * fv[i].second;
}
return norm;
}
void add_feature(const std::string& key,
double value,
common::sfv_t& ret_fv) const {
std::stringstream ss;
ss << key << "$" << value;
ret_fv.push_back(std::make_pair(ss.str(), static_cast<float>(1.0)));
}
float classifier_base::squared_norm(const common::sfv_t& fv) {
float ret = 0.f;
for (size_t i = 0; i < fv.size(); ++i) {
ret += fv[i].second * fv[i].second;
}
return ret;
}
float classifier_base::calc_margin_and_variance(
const common::sfv_t& sfv,
const string& label,
string& incorrect_label,
float& var) const {
float margin = calc_margin(sfv, label, incorrect_label);
var = 0.f;
for (size_t i = 0; i < sfv.size(); ++i) {
const string& feature = sfv[i].first;
const float val = sfv[i].second;
feature_val2_t weight_covars;
storage_->get2(feature, weight_covars);
float label_covar = 1.f;
float incorrect_label_covar = 1.f;
for (size_t j = 0; j < weight_covars.size(); ++j) {
if (weight_covars[j].first == label) {
label_covar = weight_covars[j].second.v2;
} else if (weight_covars[j].first == incorrect_label) {
incorrect_label_covar = weight_covars[j].second.v2;
}
}
var += (label_covar + incorrect_label_covar) * val * val;
}
return margin;
}
void add_feature(const std::string& key,
double value,
common::sfv_t& ret_fv) const {
ret_fv.push_back(std::make_pair(
key,
static_cast<float>(std::log(std::max(1.0, value)))));
}
void eigen_feature_mapper::rinsert(
const pair<int, float>& item,
common::sfv_t& dst) const {
if (rmap_.find(item.first) != rmap_.end()) {
dst.push_back(
make_pair((rmap_.find(item.first))->second, item.second));
}
}
void normal_herd::update(
const common::sfv_t& sfv,
float margin,
float variance,
const string& pos_label,
const string& neg_label) {
storage::storage_base* sto = get_storage();
for (common::sfv_t::const_iterator it = sfv.begin(); it != sfv.end(); ++it) {
const string& feature = it->first;
float val = it->second;
storage::feature_val2_t ret;
sto->get2(feature, ret);
storage::val2_t pos_val(0.f, 1.f);
storage::val2_t neg_val(0.f, 1.f);
ClassifierUtil::get_two(ret, pos_label, neg_label, pos_val, neg_val);
float val_covariance_pos = val * pos_val.v2;
float val_covariance_neg = val * neg_val.v2;
const float C = config_.C;
sto->set2(
feature,
pos_label,
storage::val2_t(
pos_val.v1
+ (1.f - margin) * val_covariance_pos
/ (val_covariance_pos * val + 1.f / C),
1.f
/ ((1.f / pos_val.v2) + (2 * C + C * C * variance)
* val * val)));
if (neg_label != "") {
sto->set2(
feature,
neg_label,
storage::val2_t(
neg_val.v1
- (1.f - margin) * val_covariance_neg
/ (val_covariance_neg * val + 1.f / C),
1.f
/ ((1.f / neg_val.v2) + (2 * C + C * C * variance)
* val * val)));
}
}
}
void scalar_mul_and_add(
const common::sfv_t& left,
float s,
common::sfv_t& right) {
common::sfv_t::const_iterator l = left.begin();
common::sfv_t::iterator r = right.begin();
while (l != left.end() && r != right.end()) {
if (l->first < r->first) {
std::pair<std::string, float> p = *l;
p.second *= s;
r = right.insert(r, p);
++l;
} else if (l->first > r->first) {
++r;
} else {
r->second += l->second * s;
++l;
++r;
}
}
for (; l != left.end(); ++l) {
std::pair<std::string, float> p = *l;
p.second *= s;
right.push_back(p);
}
}
vector<float> random_projection(const common::sfv_t& sfv, uint32_t hash_num) {
vector<float> proj(hash_num);
for (size_t i = 0; i < sfv.size(); ++i) {
const uint32_t seed = common::hash_util::calc_string_hash(sfv[i].first);
jubatus::util::math::random::mtrand rnd(seed);
for (uint32_t j = 0; j < hash_num; ++j) {
proj[j] += sfv[i].second * rnd.next_gaussian();
}
}
return proj;
}
void revert_feature(const common::sfv_t& fv, fv_converter::datum& data) {
for (size_t i = 0; i < fv.size(); ++i) {
std::pair<std::string, double> num_value;
std::pair<std::string, std::string> string_value;
if (revert_num_value(fv[i], num_value)) {
data.num_values_.push_back(num_value);
} else if (revert_string_value(fv[i], string_value)) {
data.string_values_.push_back(string_value);
}
}
}
common::sfv_t add(const common::sfv_t& p1, const common::sfv_t& p2) {
common::sfv_t ret;
common::sfv_t::const_iterator it1 = p1.begin();
common::sfv_t::const_iterator it2 = p2.begin();
while (it1 != p1.end() && it2 != p2.end()) {
if ((*it1).first < (*it2).first) {
ret.push_back((*it1));
++it1;
} else if ((*it1).first > (*it2).first) {
ret.push_back((*it2));
++it2;
} else {
ret.push_back(make_pair((*it1).first, (*it1).second + (*it2).second));
++it1;
++it2;
}
}
for (; it1 != p1.end(); ++it1) {
ret.push_back((*it1));
}
for (; it2 != p2.end(); ++it2) {
ret.push_back((*it2));
}
return ret;
}
double dist(const common::sfv_t& p1, const common::sfv_t& p2) {
double ret = 0;
common::sfv_t::const_iterator it1 = p1.begin();
common::sfv_t::const_iterator it2 = p2.begin();
while (it1 != p1.end() && it2 != p2.end()) {
int cmp = strcmp(it1->first.c_str(), it2->first.c_str());
if (cmp < 0) {
ret += it1->second * it1->second;
++it1;
} else if (cmp > 0) {
ret += it2->second * it2->second;
++it2;
} else {
ret += (it1->second - it2->second) * (it1->second - it2->second);
++it1;
++it2;
}
}
for (; it1 != p1.end(); ++it1) {
ret += std::pow(it1->second, 2);
}
for (; it2 != p2.end(); ++it2) {
ret += std::pow(it2->second, 2);
}
return std::sqrt(ret);
}
float recommender_base::calc_similality(common::sfv_t& q1, common::sfv_t& q2) {
float q1_norm = calc_l2norm(q1);
float q2_norm = calc_l2norm(q2);
if (q1_norm == 0.f || q2_norm == 0.f) {
return 0.f;
}
sort(q1.begin(), q1.end());
sort(q2.begin(), q2.end());
size_t i1 = 0;
size_t i2 = 0;
float ret = 0.f;
while (i1 < q1.size() && i2 < q2.size()) {
const string& ind1 = q1[i1].first;
const string& ind2 = q2[i2].first;
if (ind1 < ind2) {
++i1;
} else if (ind1 > ind2) {
++i2;
} else {
ret += q1[i1].second * q2[i2].second;
++i1;
++i2;
}
}
return ret / q1_norm / q2_norm;
}
void weight_manager::get_weight(common::sfv_t& fv) const {
for (common::sfv_t::iterator it = fv.begin(); it != fv.end(); ++it) {
double global_weight = get_global_weight(it->first);
it->second = static_cast<float>(it->second * global_weight);
}
fv.erase(remove_if(fv.begin(), fv.end(), is_zero()), fv.end());
}
void arow::update(
const common::sfv_t& sfv,
double alpha,
double beta) {
util::concurrent::scoped_wlock lk(storage_->get_lock());
for (common::sfv_t::const_iterator it = sfv.begin(); it != sfv.end(); ++it) {
const std::string& feature = it->first;
double val = it->second;
storage::feature_val2_t val2;
storage_->get2_nolock(feature, val2);
storage::val2_t current_val(0.0, 1.0);
if (val2.size() > 0) {
current_val = val2[0].second;
}
storage_->set2_nolock(
feature,
"+",
storage::val2_t(current_val.v1 + alpha * current_val.v2 * val,
current_val.v2 - beta * current_val.v2 * current_val.v2* val * val));
}
}
vector<float> euclid_lsh::calculate_lsh(const common::sfv_t& query) const {
vector<float> hash(mixable_storage_->get_model()->all_lsh_num());
for (size_t i = 0; i < query.size(); ++i) {
const uint32_t seed = common::hash_util::calc_string_hash(query[i].first);
const vector<float> proj = get_projection(seed);
for (size_t j = 0; j < hash.size(); ++j) {
hash[j] += query[i].second * proj[j];
}
}
for (size_t j = 0; j < hash.size(); ++j) {
hash[j] /= bin_width_;
}
return hash;
}
void local_storage::bulk_update(
const common::sfv_t& sfv,
float step_width,
const string& inc_class,
const string& dec_class) {
uint64_t inc_id = class2id_.get_id(inc_class);
typedef common::sfv_t::const_iterator iter_t;
if (dec_class != "") {
uint64_t dec_id = class2id_.get_id(dec_class);
for (iter_t it = sfv.begin(); it != sfv.end(); ++it) {
float val = it->second * step_width;
id_feature_val3_t& feature_row = tbl_[it->first];
feature_row[inc_id].v1 += val;
feature_row[dec_id].v1 -= val;
}
} else {
for (iter_t it = sfv.begin(); it != sfv.end(); ++it) {
float val = it->second * step_width;
id_feature_val3_t& feature_row = tbl_[it->first];
feature_row[inc_id].v1 += val;
}
}
}
