Can You Solve These Data Science Usecases?

Usecases:
1. Product Recommendation
2. Facil Recognition And Product Recommendation
3. Logisstic Procurement Using ML
4. AC Control ussing Outide Temperature
---------------------------------------------------------------------------------------------------------
All Playlist In My channel

---------------------------------------------------------------------------------------------------------------
Please donate if you want to support the channel through GPay UPID,
-------------------------------------------------------------------------------------------------------------
Please join as a member in my channel to get additional benefits like materials in Data Science, live streaming for Members and many more
-----------------------------------------------------------------------------------------------------------
Please do subscribe my other channel too
---------------------------------------------------------------------------------------------------------
Connect with me here:

[Sammy_ai]

For the use case 1 (recommendation system).

First we would need a lot of demographic information about the customers (already existing) . Do a deep dive analysis on triggers or the type of insurance they fall into. A groupby would be a nice way to check,

Secondly, just the same way an RFM analysis is done in the financial sector, check the recency, frequency and product/insurance groups you have created in the first steps. Try binning them ranking from 1 to 5 or n (depending on the distribution of the data. )

(Do a little visualization to see if the behavioral partner is captured. )
Thirdly, running a simple unsupervised clustering either (k-means or GMM initialized with k-means ) taking all precursor using the previous classification with their demographic information as parameters .

Now the recommendation system would be making recommendations to people who fall in the same final cluster.

I think this should solve the first use case. For model evaluation use the accuracy, but try reducing the false positives as that is much more dangerous business wise than false negatives.

[sinha]

1. Recommendation System:
When a person log in, get all the information regarding the person like what product he/she likes, what does the person prefer, get the browsing history of that person. Now based on the product he/she buys, we can set item-based filtering recommendation or collaborative -based filtering. Now, to tackle the cold start problem, whenever a new person comes in for the first time, show what products does he/she likes, show the tags. Once we have the data we get specific based on tag and at least show some recommendation.

2. Facial Recognition:
First we have to determine what are all the range pf products available i.e.what products are available for what part of the face. We have to check if the person is male or female. Train a model or use a pretrained model which will get the facial characteristics of the face, like if it detects a beard then recommend beard oil or a trimmer. For the female recommend cream, lotion based on the marks, spots on the face. This can be done with uploaded photos or real time camera using openCV. Similarly, like the previous to tackle cold start problem, show what products he/she is interested in, show the tags, get the history etc.

3. Logistic Procurement:
The main objective here I believe is to show whether or not we need more material or less material for the product. This is a kind of optimization problem where we are minimizing the raw materials needed. I have 2 approach for this-
a. First one is model based. First, we have to see what different types of juice are there to make. Based on that we have to see the material we need to build the product. We have get all the characteristics/features of the juice which contributes in using raw materials more or less. We can build a model based on those and either predict whether it will use more or less material (binary classification task) or predict what is amount of material the product is going to use (linear regression task).
b. The second way is based on heuristic approach. If the product use different material to to be build i.e. let say some amount of each 'n' material to make a product. Now to minimize the amount needed for each material we can use genetic algorithm to get that or other optimization algorithm to do that.Though this not so efficient, if we have lets say 100 products each using 'n' materials type as we have to do for all.

4. Car Problem:
This use case is in kinda broad spectrum. I am no knowledge in car making but still answering it.
a. Damage Detection: A car has many parts inside and outside the body. We can separate damage detection in terms of "analog i.e. mechanical parts and digital i.e. in software parts".
i. For the damage detection in mechanical parts like brakes, engine, tires etc. we can build a model which can detect damage and do segmentation of that part. For this to happen we need to install camera inside the engine hood and underneath the car. A damage can be detected not only by images but with sensors too. The car software can analyze the data to get a warning like this parts seems fishy, kindly check it out.
ii. For the damage in the software parts, show warning about the software parts.
To make it work both side needs to work together.

b. AC control: First off sense the outside temperature via sensor, sense the inside of the car also, get the temperature from both sides, then don't turn on the AC immediately just show a prompt on the screen whether or not turn on the AC. Let's keep the action on the user part. Imagine if someone has cold inside the car and suddenly it turn on the AC, good luck. Or imagine the person doesn't want the AC, he/she can pull down the windows.

[vaishupriya]

For 1st use case..we can first do clustering and segment customers into different groups and then we could apply multi-class classification based on previous data on each segment and built recommendation system based on classification results

[datahat]

For the very first use case on recommendation system. Since a rule based system was already in place, it is safe to assume that we have a dataset comprising of user records and the various policies purchased by them. Further down, if we consider that we log in the click-through rate of the recommendations as well the other policies checked in by users, collaborative recommender system could be built. Additionally, for users who already have a history of purchases, a combination of collaborative and content-based recommendation would be very useful. On the evaluation front, instead of measuring the accuracy of the recommendation, top-k accuracy would be a more suitable choice.

Some scenarios could fall into place: The company would miss out on recommending certain policies due to a lower rank score on the recommendation system (probably, historically customers did not purchase certain policy types). Instead of merely recommending the policies based on applicability, certain attribute based recommendation such as how convenient would a certain policy be for an individual as compared to the previous policies, should be combined along with other policies.

Most important thing about policies, is to maximize the revenue on the part of the insurance company as well as providing convenience and hassle free-service to the policy holders.

[PiyushSingh-uzyc]

For 2nd use case it is : Multivariate timeseries forecasting having past covariates and future covariates. We can check Darts library for this.

[ojaskulkarni]

1st use case: Recommendation system for insurance company
Sol:
User features that can be considered => age, salary, marital status, family size, dependent parents, current/past illnesses, place of residence, demographics of the area (like population, literacy, etc) which can be found through reliable online resources.
Based on user features & features of the insurance plan we can use a content based filtering recommendation system.

2nd use case: Forecast pre-pocessed food product raw material pocurement
(like Frooti or any juice)
Sol:
Features:
type of raw material,
availability/price of raw material across the year,
cost involved in storing the raw material in inventory
demand over the pas years.

Challenges:
inavailability of raw material due to rainfall or any other natural causes
transportation problems like material handling.
sudden fall in demand due to factors like defamation, competition, etc.

Model:
Use linear regression to forecast demand from past year data.
with predictor being year, population of the area and response being the predicted demand (no of products).
requirement for all the raw materials = (qty of raw material req/product) x (demand)
The procured mangoes may also have some defective pieces so we need to consider 1-2% margin over that.
Can perform RFM analysis to Recency, Frequency and Monetary to identify key demographics

3rd use case: Recommend cosmetic products based on facial features
Sol:
We can either
Use DeepFace model to extract facial features and use those features to recommend the cosmetic using decision tree classifier or content based filtering recommendation system
or
Directly feed the model with Face data use it to predict the product from available products by treating it as a classification problem.

Possible Challenges:
Bad images,
falsely identified marks on face,
sometimes the gender may also be incorrectly identified

4th use case: Maintenance alarm and AC temp adjusting model
Sol:
Take input from sensors in the car located in various components.
Use time-series analysis to determine failure/maintenance criteria set by the standards.
Combine all the data and using models based on RNN/LSTM predict the maintenance due date from current date.
Convert the time into distance-based average daily kms and inform the owner 5000 kms prior to maintenance due kms.

We'll need to consider parameters like comfort, battery or fuel usage, slope which the car is climbing, current speed of the car
to adjust AC temp. Like when climbing a slope we must turn off the AC or when the temp outside is very high we can go to lower temp but that would also depend on fuel availability.
Considering all such factors we can first try to find if the problem can be solved using linear/polynomial regression or we would need an ANN(which would also involve additional hardware in the car). So that choice again depends on car-to-car.

[datahat]

The use case of notifying the owner about potential servicing required and adjusting the internal temperature could be handled separately as: part-1--> anomaly detection in the mechanics of the engine, braking system, steering accuracy, suspension

part-2--> regulating the temperature is more of an optimizing problem, where a thermostat installed on the outside of the vehicle could be used to send signals for adjusting the temperature of the Air conditioner in the vehicle.

[mohitdubey]

for Second use case
our objective is to recommend products to specify person that can buy or willing to buy . there can more than 3 ways
way to solve 1.. step1.use person camera and to get facial features of person
step 2 use ml model to predict products prioritise
step 3 show top 3-5 priority products
way to solve 2..
using big data and social media data of our customers
train ANN or RNN and pron of this approach is the this way become better and better as increasing number of customers data

[sruthirammohan]

For the first use case (Recommendation system)

Assuming we have the details about user age, user salary, user family details, history of insurance purchase, history of renewing policies.

1. as told by you, we can use age as a factor to recommend.
2. if the user as parents who are very old, we can recommend any old-age policies we have.
3. we can recommend policies based on customer's health history.
4. based on the number of years customer is linked with us, he/she is likely to stay with us because of all the trust we have earned.
5. customer who has to support a big family but earns less

[muneebanwar]

For the second problem the actual problem here is to minimize inventory costs because if you overestimate the materials needed you need extra space and the money used on the extra inventory could have been better used elsewhere this is pretty much a demand forecasting problem from operations research I think there are alot of models for this already I remember reading about it in a book

[Mahesh]

First usecase...

Insurance recommendation system.
Basic details we get is age, sex, demography etc etc..
First we consider it is as cold start problem and we can recommend product which are similar to age interval and demographic feature etc.

After that we can track user serach and using that searches we can recommend user by using user user similarity or item item similarity.

Also we can cluster users and identify the nature and best recommendation

[atheeq]

For first use case By Using Content Based Filtering in recommended systems we can recommend anything

[mallikamehta]

for second project we can either use openCV or cascade classifier and product recomender by tensorflow or forcasting as well

[MANOJKUMAR-ccpu]

Really nice explanation regarding use cases... I expect some real world use cases related to projects... 🙏💐I hope you will upload soon thank you sir 🙏

[jinkazama-dujx]

The last use case we use Forecasting techniques plus some other algos

[karthikbhandary]

The usecase with raw products: My idea is that if we collect the data which basically impacts the production of the raw materials, We can feed it the our model to predict the amount of raw materials getting produced with give or take 5-7% error only. (We obviouly will provide the target variable as well. Also, if this is a wrong way please just tell me because I am a beginner in this!)

[adityanaranje]

For 1st use case we may use collaborative filtering

[shravananantharamakrishnan]

For the 3rd facial products use case
The face_recognition library in python enables us to extract 128d encoding of the faces based on triplet loss function
So this can be used to extract a 128d vector and since the predicted value in this case should belong to multiple classes, I will use a one layer neural network with sigmoid activation function as softmax can be used only if the predicted value belongs to only one class
False negatives will be like classifying bearded person as a non bearded person and recommending beard oil
False positives will be like classifying non beared person as a bearded person and recommending trimmers

[chiluverudivya]

Can you make video on automated roster creation with usecase

[piyush-ymuw]

For the 1st case
We use gender and age parameter and apply knn or svm