diff --git a/assign2.R b/assign2.R new file mode 100644 index 0000000000000000000000000000000000000000..66530665b367d5d39e731e29089c57506b06391a --- /dev/null +++ b/assign2.R @@ -0,0 +1,262 @@ +rm(list=ls()) +require(lubridate) +suppressWarnings(require(data.table)) +require(stringr) +require(geohash) +library(raster) +require(dplyr) +library(sp) +library(broom) +library(httr) +library(rgdal) +require(sparklyr) +require(rJava) +require(readr) + + +first = TRUE +n = 0 +for(i in 1:12){ + if(first == TRUE){ + taxi_jul <- fread("~/yellow_tripdata_2015-07.csv", + nrows=1000000) + taxi_jul <- taxi_jul %>% filter(weekdays(ymd_hms(tpep_pickup_datetime, tz = "America/New_York")) == "Wednesday") + n1 = nrow(taxi_jul) + + first <- FALSE + }else{ + taxi_bind1 <- fread("~/yellow_tripdata_2015-07.csv", + nrows=1000000, skip=n1+1) + names(taxi_bind1) <- names(taxi_jul) + + taxi_bind1 <- taxi_bind1 %>% filter(weekdays(ymd_hms(tpep_pickup_datetime, tz = "America/New_York")) == "Wednesday") + + taxi_jul <- bind_rows(taxi_jul, taxi_bind1) + + n1 = nrow(taxi_jul) + + } +} + +taxi <- taxi_jul +rm(taxi_jul) + +hm_pickup <- (lubridate::hour(taxi$tpep_pickup_datetime)*60)+as.numeric(lubridate::minute(taxi$tpep_pickup_datetime)) + +interval_5min <- vector() +for(j in seq_len(nrow(taxi))){ + for(i in seq(0, 2350, by = 5)){ + if(between(hm_pickup[j], i, i+5)){ + interval_5min[j] = paste0("(", i, " - ", i+5, "]") + } + } +} + +taxi$interval_5min <- interval_5min + +#convert dropoff times to ordered factor with monday as 1, sunday as 7 +dropoff_weekday_num <- weekdays(ymd_hms(taxi$tpep_dropoff_datetime, tz = "America/New_York")) + +dropoff_weekday_num <- factor(dropoff_weekday_num, + levels = c("Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"), + ordered = TRUE) +pickup_weekday_num <- weekdays(ymd_hms(taxi$tpep_pickup_datetime, tz = "America/New_York")) + +pickup_weekday_num <- factor(pickup_weekday_num, + levels = c("Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"), + ordered = TRUE) + + +#convert time to seconds and then make cosine and sine versions for smooth transitions between time intervals +taxi <- taxi %>% dplyr::mutate(time_hour = lubridate::hour(taxi$tpep_pickup_datetime), + time_num = (((lubridate::hour(taxi$tpep_pickup_datetime)*60*60)+ + (lubridate::minute(taxi$tpep_pickup_datetime)*60)+ + lubridate::second(taxi$tpep_pickup_datetime))/(60*60*24)), + time_num_cosine = cos(time_num*2*pi), + time_num_sine = sin(time_num*2*pi)) + +#add a feature for weekday as a numeric value - 1 for monday, 7 for sunday +taxi <- taxi %>% dplyr::mutate(weekday = pickup_weekday_num, + weekday_num = (as.numeric(weekday)+time_num)/7, + weekday_num_cosine = cos(weekday_num*2*pi), + weekday_num_sine = sin(weekday_num*2*pi)) + +#time difference between pickup and dropoff in seconds +taxi <- taxi %>% dplyr::mutate(time_diff_secs = + ((lubridate::hour(taxi$tpep_dropoff_datetime)*60*60)+ + (lubridate::minute(taxi$tpep_dropoff_datetime)*60)+ + lubridate::second(taxi$tpep_dropoff_datetime)) - + ((lubridate::hour(taxi$tpep_pickup_datetime)*60*60)+ + (lubridate::minute(taxi$tpep_pickup_datetime)*60)+ + lubridate::second(taxi$tpep_pickup_datetime))) + +#create week as numeric value +taxi <- taxi %>% dplyr::mutate(weekofyear = lubridate::week(tpep_pickup_datetime), + weekofyear_num = (weekofyear+weekday_num)/53, + weekofyear_num_cosine = cos(weekofyear_num*2*pi), + weekofyear_num_sine = sin(weekofyear_num*2*pi)) + + +#remove impossible and unlikely values +taxi <- taxi %>% dplyr::filter(time_diff_secs > 0 & time_diff_secs/60 < 300) +#convert time difference feature to be between 0 and 1 (scaled by max) +taxi <- taxi %>% dplyr::mutate(timediffsec_num = time_diff_secs/max(time_diff_secs)) + + +#miles to kms +taxi$trip_distance <- taxi$trip_distance*1.609344 + +#outlier for trip distance, just remove as don't know how else to deal with it +taxi <- taxi %>% dplyr::filter(trip_distance > 0 & trip_distance < 150) +#remove longitude and latitude equal to 0 +taxi <- taxi %>% dplyr::filter(pickup_longitude != 0 | pickup_latitude != 0) + + +taxi <- taxi %>% dplyr::filter(trip_distance > 0) +taxi <- taxi %>% dplyr::filter(total_amount < 200) + + +#data features inspired by https://sdaulton.github.io/TaxiPrediction/ +#specifically geohashing, find out this is best in analysis below comparing to boroughs and neighbourhoods +#Also numeric values between 0 and 1 for time of day, day of week, week of year +#And extra features converting these scaled numeric values functions of sine and cosine +#to make for a more smooth transition between time periods + +taxi$pickup_geohash <-gh_encode(taxi$pickup_latitude, taxi$pickup_longitude, 5) + +#scale trip distance to be between 0 and 1 +taxi$trip_distance_num <- taxi$trip_distance/max(taxi$trip_distance) +taxi$RatecodeID <- as.factor(taxi$RatecodeID) +taxi$passenger_count <- as.factor(taxi$passenger_count) + +#scale total amount to be between 0 and 1 +taxi$total_amount_num <- taxi$total_amount/max(taxi$total_amount) + + +#59 unique geohashes for pickups +taxi %>% count(pickup_geohash) + +r <- GET('http://data.beta.nyc//dataset/0ff93d2d-90ba-457c-9f7e-39e47bf2ac5f/resource/35dd04fb-81b3-479b-a074-a27a37888ce7/download/d085e2f8d0b54d4590b1e7d1f35594c1pediacitiesnycneighborhoods.geojson') + +nyc_neighborhoods <- readOGR(content(r,'text'), 'OGRGeoJSON', verbose = F) + +taxi <- as.data.frame(taxi) +taxi_spdf <- SpatialPointsDataFrame(taxi[,c('pickup_longitude', 'pickup_latitude')], + proj4string = CRS("+proj=longlat +datum=WGS84 +no_defs +ellps=WGS84 +towgs84=0,0,0"), + data=taxi) +taxi_spdf <- over(taxi_spdf, nyc_neighborhoods[,c('neighborhood', 'borough')]) +taxi$neighbourhood <- taxi_spdf$neighborhood +taxi$borough <- taxi_spdf$borough + +glimpse(taxi) +taxi %>% group_by(neighbourhood, interval_5min) %>% count() %>% arrange(desc(n)) %>% + group_by(neighbourhood) %>% summarise(n = sum(n)) %>% arrange(desc(n)) + +#incredible that there is upward of 500 pickups in 5 minute intervals on a wednesday in Midtown in NYC +taxi %>% filter(neighbourhood %in% c("Midtown", "Upper East Side", "Chelsea", "Upper West Side", + "Hell's Kitchen")) %>% count(interval_5min) %>% + ggplot() + geom_col(aes(x=interval_5min, y=n)) + + theme(axis.text.x = element_text(angle = 55, hjust = 1)) + ggtitle("Pickup density for Midtown on Wednesday's in July of 2015") + +tt <- taxi %>% filter(neighbourhood %in% c("Midtown", "Upper East Side")) %>% group_by(neighbourhood, interval_5min) %>% + count() %>% ungroup(taxi) +taxi2 <- taxi %>% filter(neighbourhood %in% c("Midtown", "Upper East Side")) +taxi3 <- taxi %>% filter(neighbourhood %in% c("Midtown", "Upper East Side")) +dim(taxi2) +sort(tt$n, decreasing = T) +pickup_n <- vector() +tt$neighbourhood <- as.character(tt$neighbourhood) +tt_mid <- tt %>% filter(neighbourhood == "Midtown") +names(tt_mid)[3] <- "n1" +tt_ues <- tt %>% filter(neighbourhood == "Upper East Side") +names(tt_ues)[2:3] <- c("neighbourhood2", "n2") +tt_ues <- tt_ues %>% dplyr::select(neighbourhood2, n2) +tt <- cbind(tt_mid, tt_ues) +head(tt) +for(i in 1:nrow(taxi2)){ + for(j in 1:length(unique(tt$interval_5min))){ + if(taxi2$interval_5min[i] == tt$interval_5min[j]){ + if(taxi2$neighbourhood[i] == tt$neighbourhood[j]){ + pickup_n[i] = tt$n1[j] + }else{ + pickup_n[i] = tt$n2[j] + } + } + } +} +taxi2$pickup_n <- pickup_n +glimpse(taxi2) +taxi2 %>% filter(neighbourhood == "Midtown") %>% count(interval_5min) %>% + ggplot() + geom_col(aes(x=reorder(interval_5min, n), y=n)) + + theme(axis.text.x=element_blank(), axis.ticks.x=element_blank()) + xlab("Time")+ggtitle("Midtown") +taxi2 %>% filter(neighbourhood == "Upper East Side") %>% count(interval_5min) %>% + ggplot() + geom_col(aes(x=reorder(interval_5min, n), y=n)) + theme(axis.text.x=element_blank(), + axis.ticks.x=element_blank()) +ggtitle("Upper East Side") + xlab("Time") + +taxi2 %>% filter(neighbourhood == "Midtown") %>% count(interval_5min) %>% mutate(sd = sd(n)) +taxi2 %>% filter(neighbourhood == "Upper East Side") %>% count(interval_5min) %>% mutate(sd = sd(n)) + +glimpse(taxi2) +taxi2$VendorID <- as.factor(taxi2$VendorID) +taxi2$payment_type <- as.factor(taxi2$payment_type) + +taxi2 <- taxi2 %>% dplyr::select(VendorID, passenger_count, RatecodeID, payment_type, time_num, time_num_cosine, time_num_sine, + weekday_num, weekday_num_cosine, weekday_num_sine, timediffsec_num, trip_distance_num, + total_amount_num, neighbourhood, interval_5min, pickup_n) + +taxi2 %>% filter(neighbourhood == "Midtown") %>% group_by(interval_5min) %>% + summarise(tan = median(total_amount_num), tdn = median(trip_distance_num), tdsn = median(timediffsec_num), + woyns = mean(weekofyear_num_sine), woync = mean(weekofyear_num_cosine), woyn = median(weekofyear_num), + tns = mean(time_num_sine), tnc = mean(time_num_cosine), tn = median(time_num)) + +Y <- taxi %>% filter(neighbourhood == "Midtown") %>% count(pickup_halfhr_cat) %>% dplyr::select(n) +taxi <- taxi %>% dplyr::mutate(Y = Y$n) + +taxi2 %>% filter(neighbourhood == "Midtown") %>% + ggplot() + geom_col(aes(x=interval_5min, y=pickup_n)) +write.csv(taxi2, "D:/R/TaxiAssignment/taxi_ml_format.csv") + + +#######SPARK +cluster_url <- paste0("spark://", system("hostname -i", intern = TRUE), ":7077") +library(sparklyr) +library(readr) +library(dplyr) +sc <- spark_connect(master = cluster_url) +spark_read_csv(sc, "taxi_ml_format.csv", path = "") +taxi_ml_format <- read_csv("taxi_ml_format.csv") +taxi_ml_format <- taxi_ml_format %>% select(-X1) +X_data <- taxi_ml_format + + +X_data$VendorID <- as.factor(X_data$VendorID) +X_data$RatecodeID <- as.factor(X_data$RatecodeID) +X_data$payment_type <- as.factor(X_data$payment_type) +X_data$passenger_count <- as.factor(X_data$passenger_count) + +X_tbl <- copy_to(sc, X_data, "x_data", overwrite=T) + +partitions <- X_tbl %>% + sdf_partition(training = 0.75, test = 0.25, seed = 1099) + +taxi_training <- partitions$training +taxi_test <- partitions$test + +#baseline +lm <- ml_linear_regression(taxi_training, pickup_n~.) + +rf_model <- taxi_training %>% + ml_random_forest(pickup_n ~ ., type = "regression") + +pred_lm <- sdf_predict(taxi_test, lm) +pred_rf <- sdf_predict(taxi_test, rf_model) + +ml_regression_evaluator(pred_lm, label_col = "pickup_n", prediction_col = "prediction", metric_name="r2") +ml_regression_evaluator(pred_rf, label_col = "pickup_n", prediction_col = "prediction", metric_name="r2") +ml_regression_evaluator(pred_lm, label_col = "pickup_n", prediction_col = "prediction", metric_name="rmse") +ml_regression_evaluator(pred_rf, label_col = "pickup_n", prediction_col = "prediction", metric_name="rmse") + +imp <- sparklyr::ml_feature_importances(rf_model) + + +spark_disconnect(sc) \ No newline at end of file