Google Maps Restaurant Type Classification

 

Description:

A large percentage of restaurants in Google Maps are either labelled null or generic as seen in Google Maps/Places which need to be tagged to correct cuisine/type of restaurant.

rest_type	cnt_per_group	perc_of_count_total
restaurant	9992	16.69925628812568
mexican	4073	6.807052728336258
pizza	2862	4.783153672599649
null	2401	4.0127015960558206
fast food	2328	3.89069942341439
sandwich	2235	3.735271997994485
chinese	2145	3.584858360491351
american	1930	3.225536893122754
italian	1693	2.829447647697836
coffee	1693	2.829447647697836
seafood	995	1.6629063257290881
japanese	970	1.621124759755996
thai	922	1.5409041530876577
bar & grill	921	1.539232890448734
sushi	887	1.4824099607253278
hamburger	848	1.4172307178073034
chicken	828	1.3838054650288292
cafe	717	1.1982953121082978
bar	650	1.0863207153004095
indian	614	1.026155260299156

To resolve this issue, we use a NLP based classification model on known restaurant types as labels and reviews as input. Reviews often contain information about the food items as well the cuisine information that can be useful to tag the restaurants with the correct cuisines.

 Cleaning Restaurant Types:

The restaurant types are combined for similar cuisines as seen below

df_attrs_clean = df_rest_attrs.withColumn('rest_type', 
              F.when(F.lower(F.col('rest_type')).rlike('taco|mexican|burrito|mex'), F.lit('mexican'))\
              .when(F.lower(F.col('rest_type')).rlike('bar|pub|brewery|beer|gastropub|brasserie|bistro'), F.lit('bar'))\
              .when(F.lower(F.col('rest_type')).rlike('pizza'), F.lit('pizza'))\
              .when(F.lower(F.col('rest_type')).rlike('irish|fish & chips|fish and chips'), F.lit('irish'))\
              .when(F.lower(F.col('rest_type')).rlike('indian|pakistan|nepalese'), F.lit('indian'))\
              .when(F.lower(F.col('rest_type')).rlike('chinese|dim|sichuan|noodle|mandarin|shang|hong|\
                                                      |hot pot|餐馆|cantonese|dumpling'), F.lit('chinese'))\
              .when(F.lower(F.col('rest_type')).rlike('american|us|hot dog|diner|southern|cajun'), F.lit('american'))\
              .when(F.lower(F.col('rest_type')).rlike('ramen|japanese|izakaya'), F.lit('japanese'))\
              .when(F.lower(F.col('rest_type')).rlike('sushi'), F.lit('sushi'))\
              .when(F.lower(F.col('rest_type')).rlike('steak'), F.lit('steak'))\
              .when(F.lower(F.col('rest_type')).rlike('thai|cambodia'), F.lit('thai'))\
              .when(F.lower(F.col('rest_type')).rlike('chicken'), F.lit('fast food'))\
              .when(F.lower(F.col('rest_type')).rlike('vietnamese|pho'), F.lit('vietnamese'))\
              .when(F.lower(F.col('rest_type')).rlike('italian|pasta'), F.lit('italian'))\
                                          .when(F.lower(F.col('rest_type')).rlike('portuguese|salvadoran|peru|\
                                                    |cuban|brazilian|honduran|latin|guatemalan|ecuadorian|\
                                                      |argentinian|venezuelan|colombian|dominican|nicaraguan')\
                    , F.lit('latin american'))\
                                     
              
              .when(F.lower(F.col('rest_type')).rlike('greek|gyro|falafel|middle eastern|afghani|persian|\
                                                  |halal|kebab|mediterranean|middle eastern|lebanese'), F.lit('mediterranean'))\
              .when(F.lower(F.col('rest_type')).rlike('korean|한식당|음식점'), F.lit('korean'))\
              .when(F.lower(F.col('rest_type')).rlike('french|crêperie|crepe'), F.lit('french'))\
              .when(F.lower(F.col('rest_type')).rlike('lounge|bistro'), F.lit('bistro'))\
              .when(F.lower(F.col('rest_type')).rlike('burger|sandwich'), F.lit('burger'))\
              .when(F.lower(F.col('rest_type')).rlike('brunch|breakfast|pancake'), F.lit('brunch'))\
              .when(F.lower(F.col('rest_type')).rlike('coffee|cafe|bake|bakery|donut|bagel'), F.lit('cafe&bakery'))\
              .when(F.lower(F.col('rest_type')).rlike('southern|soul food'), F.lit('southern'))\
              .when(F.lower(F.col('rest_type')).rlike('spanish|tapas'), F.lit('spanish'))\
              .when(F.lower(F.col('rest_type')).rlike('asian'), F.lit('asian'))\
              .when(F.lower(F.col('rest_type')).rlike('european|german|polish|belgian|british|swedish|irish'),   F.lit('european'))\
              .when(F.lower(F.col('rest_type')).rlike('african'), F.lit('african'))\
              .when(F.lower(F.col('rest_type')).rlike('vegan|vegetarian'), F.lit('vegetarian'))\
              .when(F.lower(F.col('rest_type')).rlike('stand|venue|club|mall|alley|market|store|association|\
                                                      |station|juice|ice cream|center|theater|cater|court|fuel|\
                                                      |tobacco|arcade|producer|winery|yogurt|hall|school|grovery|service|\
                                                      |apartment|agency|organization|atm|estate|office|casino|\
                                                      |company|consultant|gift|deli'), F.lit('remove')) \
              .otherwise(F.lit(F.col('rest_type'))))

 Model:

 Sequential Model with Word Embeddings

def generate_model_lstm(total_words, max_len, EMBED_DIM = 512):
    model = Sequential()
    model.add(Embedding(input_dim=total_words, 
                               output_dim=EMBED_DIM, 
                               input_length=max_len))
    model.add(LSTM(128, return_sequences= True))
    model.add(GlobalMaxPool1D())
    model.add(Dense(128, activation='relu'))
    model.add(Dropout(0.3))
    model.add(Dense(15, activation='softmax'))
    model.compile(optimizer='adam',
                  loss='categorical_crossentropy',
                  metrics=['accuracy'])
    return model

Structure

Model Results:

  

Accuracy: 83.46%

• Model accuracy > 80%
• Many restaurants are tagged as American
• Lower classification accuracy (< 70%) for Bar, Japanese and Seafood restaurants
• Mexican , Thai, Chinese, Pizza and India restaurants have accuracies > 90%