Analyzing rtweet data with kerasformula

Transcription

1 Analyzing rtweet data with kerasformula Pete Mohanty BARUG March 2018 hosted by Google

2 Overview Now on CRAN, kerasformula package offers a high-level interface for the R interface to Keras. install.packages("kerasformula") library(kerasformula) install_keras() # or install_keras(tensorflow = "gpu")

3 Overview Now on CRAN, kerasformula package offers a high-level interface for the R interface to Keras. install.packages("kerasformula") library(kerasformula) install_keras() # or install_keras(tensorflow = "gpu") kerasformula enables users to train and test neural nets via Keras starting with raw data in as little as one line of code.

4 Overview Now on CRAN, kerasformula package offers a high-level interface for the R interface to Keras. install.packages("kerasformula") library(kerasformula) install_keras() # or install_keras(tensorflow = "gpu") kerasformula enables users to train and test neural nets via Keras starting with raw data in as little as one line of code. This talk introduces the regression-style interface using data gathered via rtweet (developed

5 kms() kerasformula s main function Most machine learning demos assume homogenous data (e.g., pixels for digit recognition) which can make coding to fit heterogeneous data challenging. kms() takes advantage of R formulas to smooth this process.

6 kms() kerasformula s main function Most machine learning demos assume homogenous data (e.g., pixels for digit recognition) which can make coding to fit heterogeneous data challenging. kms() takes advantage of R formulas to smooth this process. kms builds dense neural nets and, after fitting them, returns a single object with predictions, measures of fit, and details about the function call.

7 kms() kerasformula s main function Most machine learning demos assume homogenous data (e.g., pixels for digit recognition) which can make coding to fit heterogeneous data challenging. kms() takes advantage of R formulas to smooth this process. kms builds dense neural nets and, after fitting them, returns a single object with predictions, measures of fit, and details about the function call. kms accepts many parameters found in keras like loss and activation functions.

8 kms() kerasformula s main function Most machine learning demos assume homogenous data (e.g., pixels for digit recognition) which can make coding to fit heterogeneous data challenging. kms() takes advantage of R formulas to smooth this process. kms builds dense neural nets and, after fitting them, returns a single object with predictions, measures of fit, and details about the function call. kms accepts many parameters found in keras like loss and activation functions. kms also accepts compiled keras_model_sequential objects allowing for even further customization.

9 The Data Let s look at #rstats tweets (excluding retweets) for a six-day period ending January 24, 2018 at 10:40. rstats <- search_tweets("#rstats", n = 10000, include_rts = FALSE) dim(rstats) [1] cor(rstats$favorite_count, rstats$retweet_count, method="spearman") [1]

10 Skewed against Popularity

11 Goal - predict popularity of each tweet rstats$popularity <- rstats$retweet_count + rstats$favorite_count Dependent variable will be discretized into approximate levels.

12 Getting the most out of formulas breaks <- c(-1, 0, 1, 10, 100, 1000, 10000) popularity <- kms(cut(popularity, breaks) ~ screen_name + source + n(hashtags) + n(mentions_screen_name) + n(urls_url) + nchar(text) + grepl('photo', media_type) + weekdays(created_at) + format(created_at, '%H'), rstats)

13 Model History

14 Confusion Matrix for First Model popularity$confusion (-1,0] (0,1] (1,10] (10,100] (100, 1000] (1000, (-1,0] (0,1] (1,10] (10,100] (100, 1000] (1000, 10000]

15 Storing the Input Formula... form <- "cut(popularity, breaks) ~ n(hashtags) + n(mentions_screen_n nchar(text) + screen_name + source + grepl('photo', media_type) + weekdays(created_at) + format(created_at, '%H')"

16 Adding Mentions Which, if any, other Twitter users are mentioned in a Tweet comes buried in variable-length list corresponding to each tweet. grepl(" ", mentions_user_id) Adding to the formula allows kms() to construct sparse model matrices directly.

17 Adding Mentions Which, if any, other Twitter users are mentioned in a Tweet comes buried in variable-length list corresponding to each tweet. grepl(" ", mentions_user_id) Adding to the formula allows kms() to construct sparse model matrices directly. M <- unique(unlist(rstats$mentions_user_id)) M <- unique(mentions[which(table(m) > 5)]) M <- M[!is.na(M)] for(m in M) form <- paste(form, "+ grepl(", m,", mentions_user_id)") popularity <- kms(pop_input, rstats)

18

19 Customizing layers with kms() Say we want to transition more gradually from the input shape to the output. l <- list(units= c(512, 256, 128, NA), activation= c("relu", "relu", "relu", "softmax"), dropout= c(0.45, 0.4, 0.35, NA)) out <- kms(pop_input, rstats, layers = l)

20 Making predictions for new data popularity <- kms(pop_input, rstats[1:1000,]) predictions <- predict(popularity, rstats[1001:2000,]) predictions$accuracy [1] 0.579

21 Making predictions for new data popularity <- kms(pop_input, rstats[1:1000,]) predictions <- predict(popularity, rstats[1001:2000,]) predictions$accuracy [1] kerasformula s predict method will automatically handle mismatches in columns between the training and the new data.

22 Using a compiled Keras model k <- keras_model_sequential() k %>% layer_embedding(input_dim = out$p, output_dim = out$p) %>% layer_lstm(units = 512, dropout = 0.4, recurrent_dropout = 0.2) %>% layer_dense(units = 256, activation = "relu") %>% layer_dropout(0.3) %>% layer_dense(units = 8, # y's number of levels activation = 'sigmoid')

23 Using a compiled Keras model k <- keras_model_sequential() k %>% layer_embedding(input_dim = out$p, output_dim = out$p) %>% layer_lstm(units = 512, dropout = 0.4, recurrent_dropout = 0.2) %>% layer_dense(units = 256, activation = "relu") %>% layer_dropout(0.3) %>% layer_dense(units = 8, # y's number of levels activation = 'sigmoid') k %>% compile( loss = 'categorical_crossentropy', optimizer = 'rmsprop', metrics = c('accuracy') )

24 Using a compiled Keras model k <- keras_model_sequential() k %>% layer_embedding(input_dim = out$p, output_dim = out$p) %>% layer_lstm(units = 512, dropout = 0.4, recurrent_dropout = 0.2) %>% layer_dense(units = 256, activation = "relu") %>% layer_dropout(0.3) %>% layer_dense(units = 8, # y's number of levels activation = 'sigmoid') k %>% compile( loss = 'categorical_crossentropy', optimizer = 'rmsprop', metrics = c('accuracy') ) out_lstm <- kms(pop_input, rstats, k)

25 Thanks! Questions? Comments? Special thanks to Dan Falbel and JJ Allaire for helpful suggestions and Joe Rickert for organizing!! Tonight s kerasformula code can be found on the RStudio Tensorflow page or the project s github repo via:

26 Choosing a Batch Size By default, kms uses batches of 32. Suppose we were happy with our model but didn t have any particular intuition about what the size should be. Nbatch <- c(16, 32, 64) Nruns <- 4 accuracy <- matrix(nrow = Nruns, ncol = length(nbatch)) colnames(accuracy) <- paste0("nbatch_", Nbatch) est <- list() for(i in 1:Nruns){ for(j in 1:length(Nbatch)){ est[[i]] <- kms(pop_input, rstats, Nepochs = 2, batch_si accuracy[i,j] <- est[[i]][["evaluations"]][["acc"]] } } colmeans(accuracy)