Smartphone Similarity Search

What are we going to do?

In this quick demo, we will use JAI to:

• Train and deploy models into a secure and scalable production-ready environment.

• Complete a similarity search - Given a database of smartphone models and its attributes, identify which smartphones are most similar to the target.

---

Importing libraries

import pandas as pd
from jai import Jai

---

JAI Auth Key

If you don’t already have an auth key, you can get your auth key here - free forever. Also, please make sure to check your spam folder if you can’t find it in your inbox!

from jai import get_auth_key
get_auth_key(email = 'email@emailnet.com', firstName = 'JAI', lastName = 'Z')

<Response [201]>

---

Dataset quick look

This dataset contains data related to the specs of different models of smartphones. Some of the analyzed specs are, for example, dimensions, display type, model and brand. In this example, we’re going to identify which cellphones are most similar to one another based on their sspecs.

# Obtaining the data from website
file_url = 'https://myceliademo.blob.core.windows.net/smartphone-dataset/smartphones_sample_db.csv?sp=rl&st=2021-05-17T16:30:09Z&se=2025-01-18T16:30:00Z&sv=2020-02-10&sr=b&sig=6LeB2OPLM33LXiPaaQ6LvO00mt0MMfgczKtt92AJvMU%3D'
df = pd.read_csv(file_url,index_col=0)

# Show name of columns and non-null count
df.info()

<class 'pandas.core.frame.DataFrame'>
Int64Index: 1000 entries, 4439 to 2276
Data columns (total 37 columns):
 #   Column              Non-Null Count  Dtype
---  ------              --------------  -----
 0   objectId            1000 non-null   object
 1   Model               999 non-null    object
 2   Brand               999 non-null    object
 3   Network             999 non-null    object
 4   TwoG                970 non-null    object
 5   ThreeG              571 non-null    object
 6   FourG               199 non-null    object
 7   Network_Speed       574 non-null    object
 8   GPRS                996 non-null    object
 9   EDGE                997 non-null    object
 10  Announced           998 non-null    object
 11  Status              999 non-null    object
 12  Dimensions          998 non-null    object
 13  field13             69 non-null     object
 14  SIM                 999 non-null    object
 15  Display_type        999 non-null    object
 16  Display_resolution  874 non-null    object
 17  Display_size        995 non-null    object
 18  Operating_System    569 non-null    object
 19  CPU                 561 non-null    object
 20  Chipset             419 non-null    object
 21  GPU                 409 non-null    object
 22  Memory_card         999 non-null    object
 23  Internal_memory     791 non-null    object
 24  RAM                 543 non-null    object
 25  Primary_camera      861 non-null    object
 26  Secondary_camera    859 non-null    object
 27  Loud_speaker        999 non-null    object
 28  Audio_jack          996 non-null    object
 29  WLAN                998 non-null    object
 30  Bluetooth           999 non-null    object
 31  GPS                 998 non-null    object
 32  NFC                 101 non-null    object
 33  Radio               984 non-null    object
 34  USB                 902 non-null    object
 35  Sensors             551 non-null    object
 36  Battery             999 non-null    object
dtypes: object(37)
memory usage: 296.9+ KB

# Showing first 5 rows of dataframe
df.head()

	objectId	Model	Brand	Network	TwoG	ThreeG	FourG	Network_Speed	GPRS	EDGE	...	Loud_speaker	Audio_jack	WLAN	Bluetooth	GPS	NFC	Radio	USB	Sensors	Battery
4439	9X9xdNsAcI	_M32	Bird	GSM	GSM 900 / 1800	NaN	NaN	NaN	Class 10	No	...	Yes	No	No	No	No	NaN	FM radio (optional)	Yes	NaN	Removable Li-Ion 600 mAh battery
2922	uWQacAjFLE	_Galaxy J3 Pro	Samsung	GSM / CDMA / HSPA / EVDO / LTE	GSM 850 / 900 / 1800 / 1900 - SIM 1 & SIM 2	HSDPA 850 / 900 / 1900 / 2100	LTE band 1(2100)| 3(1800)| 7(2600)| 41(2500)	HSPA 42.2/5.76 Mbps LTE Cat4 150/50 Mbps	Yes	Yes	...	Yes	Yes	Wi-Fi 802.11 b/g/n| Wi-Fi Direct| hotspot	4.1| A2DP	Yes with A-GPS GLONASS BDS	Yes	FM radio| RDS| recording	microUSB 2.0| USB On-The-Go	Accelerometer| proximity	Removable Li-Ion 2600 mAh battery
4071	D60tQhWyLw	_KP501 Cookie	LG	GSM	GSM 850 / 900 / 1800 / 1900	NaN	NaN	NaN	Class 10	Class 10	...	Yes	No	No	2.1| A2DP	No	NaN	Stereo FM radio| RDS	2	Accelerometer	Removable Li-Ion 900 mAh battery
3689	JgAmFNzZGT	_Iconic Phablet	ZTE	CDMA / EVDO / LTE	CDMA 800 / 1900	CDMA2000 1xEV-DO	LTE	EV-DO Rev.A 3.1 Mbps LTE	Yes	Yes	...	Yes	Yes	Wi-Fi 802.11 b/g/n| hotspot	4.0| A2DP	Yes with A-GPS	NaN	To be confirmed	microUSB 2.0	Accelerometer| proximity	Non-removable Li-Ion 3200 mAh battery
3882	taHMJkJclc	_J501	Asus	GSM	GSM 900 / 1800 / 1900	NaN	NaN	NaN	Class 12	No	...	No	No	No	1.2| A2DP	No	NaN	FM radio	1.1	NaN	Removable Li-Ion 750 mAh battery

5 rows × 37 columns

---

Search by product attribute

Now we’re going to execute a similarity search by product attribute. To do so, we first need to insert the smartphone dataframe into jai. The method used to send data to Jai is j.setup (or j.fit; they are the same), which can then be consumed through the methods j.similar and j.predict.

# Instantiate Jai class
j = Jai()

# Using jai setup with self-supervised learning
# The raw data is inserted into jai, which turns the data into a latent vector collection that is then stored in a database

j.setup(
    name = 'smartphones',
    verbose = 2,
    data = df,
    db_type = 'SelfSupervised',
    overwrite = True
)

Insert Data: 100%|██████████| 1/1 [00:00<00:00,  3.08it/s]

Training might finish early due to early stopping criteria.

Recognized setup args:
- db_type: SelfSupervised

JAI is working: 100%|██████████|20/20 [00:10]

Setup Report:

Best model at epoch: 19 val_loss: 0.07

({0: {'Task': 'Adding new data for tabular setup',
   'Status': 'Completed',
   'Description': 'Insertion completed.',
   'Interrupted': False}},
 {'Task': 'Training Model',
  'Status': 'Job Created',
  'Description': 'Check status after some time!',
  'kwargs': {'db_type': '"SelfSupervised"'}})

---

Checking the collections available on Jai

# Verifying databases that have already been created for the used AuthKey
j.names

['california_housing', 'census', 'smartphones']

j.info
# Gets name, type and other information about each created database

	name	type	last modified	dependencies	size	embedding_dimension
2	california_housing	Supervised	2022-06-20-14h39	[]	20640	64
1	census	Supervised	2022-06-20-14h41	[]	32561	64
0	smartphones	SelfSupervised	2022-06-20-14h43	[]	1000	64

---

Search similar smartphones - by all attributes

Once the data is inserted into Jai, we can execute a similarity search to find smartphones that are most similar to the one we want. The search is based on the distance between the vector representations of the smartphones, meaning that a smartphone will be considered to be the most similar to another if the distance between them is smaller than the distance between that smartphone and any other one.

# We can execute a similarity search on a specific index, which is the index of the initial pandas dataframe
results = j.similar('smartphones', [856], top_k=5)
results
# Smaller distances between smartphones mean that they are more similar
# The distance of 0.0 is obtained when comparing the smartphone to itself

Similar: 100%|██████████| 1/1 [00:00<00:00,  2.85it/s]

[{'query_id': 856,
  'results': [{'id': 856, 'distance': 0.0},
   {'id': 1978, 'distance': 0.6144497394561768},
   {'id': 7040, 'distance': 0.7259373664855957},
   {'id': 1597, 'distance': 0.7516287565231323},
   {'id': 6476, 'distance': 0.7594107985496521}]}]

# We can also do it on the whole dataframe at once
res = j.similar('smartphones', df.index, top_k = 5)
# By doing this, we are finding the top 5 most similar smartphones to each smartphone on the initial dataframe
# Result is a list of dictionaries, where each dictionary shows the id of the top 5 results
res[0]

Similar: 100%|██████████| 1/1 [00:01<00:00,  1.84s/it]

{'query_id': 4439,
 'results': [{'id': 4439, 'distance': 0.0},
  {'id': 4490, 'distance': 0.5862273573875427},
  {'id': 8627, 'distance': 0.6433966159820557},
  {'id': 1483, 'distance': 0.7437235713005066},
  {'id': 6375, 'distance': 0.7853354215621948}]}

# Now, we use the IDs we found on the last step to locate the most similar smartphones on the initial dataframe
df.loc[pd.DataFrame(res[0]['results']).id]

	objectId	Model	Brand	Network	TwoG	ThreeG	FourG	Network_Speed	GPRS	EDGE	...	Loud_speaker	Audio_jack	WLAN	Bluetooth	GPS	NFC	Radio	USB	Sensors	Battery
4439	9X9xdNsAcI	_M32	Bird	GSM	GSM 900 / 1800	NaN	NaN	NaN	Class 10	No	...	Yes	No	No	No	No	NaN	FM radio (optional)	Yes	NaN	Removable Li-Ion 600 mAh battery
4490	vxtdVQdFMg	_M580	BenQ	GSM	GSM 900 / 1800 / 1900	NaN	NaN	NaN	Class 10	No	...	No	No	No	1.2	No	NaN	No	No	NaN	Removable Li-Ion 870 mAh battery
8627	DJy3DEvEue	_Zum	Parla	GSM	GSM 850 / 900 / 1800 / 1900	NaN	NaN	NaN	No	No	...	Yes	No	No	No	No	NaN	Stereo FM radio	miniUSB	NaN	Removable Li-Ion 700 mAh battery
1483	7VWSVUS0lt	_C30	BenQ	GSM	GSM 900 / 1800	NaN	NaN	NaN	Class 12	No	...	Yes	No	No	No	No	NaN	FM radio	1.1	NaN	Removable Li-Ion 650 mAh battery
6375	U5MelmKcdH	_S1600	Kyocera	GSM	GSM 850 / 1900	NaN	NaN	NaN	Class 10	No	...	Yes	No	No	No	No	NaN	No	microUSB	NaN	Removable Li-Ion 650 mAh battery

5 rows × 37 columns

---

Requests via REST API

import requests

url = 'https://mycelia.azure-api.net/similar/id/smartphones?id=841&top_k=5'
auth_header = {'Auth': "INSERT_YOUR_AUTH_KEY_HERE"}

response = requests.get(url, headers=auth_header)
response.json()

{'similarity': [{'query_id': 841,
   'results': [{'distance': 0.0, 'id': 841},
    {'distance': 0.5942456126213074, 'id': 856},
    {'distance': 0.7880990505218506, 'id': 3885},
    {'distance': 0.8442420959472656, 'id': 1597},
    {'distance': 0.895087480545044, 'id': 2531}]}]}