Addressing the Cold Start Problem in Streaming Recommendations

Background

Recommendation systems in online streaming are those systems which basically provides the users’ the recommendation of movies, seasons, dramas and documentaries similar to their interest. These systems store the users interest as an input and on the basis of stored interest, these systems try to recommend similar things to a number of users.

Motivation

Recommendation system is one of the system which is now considered essential in the field of E-Commerce, online streaming, Online gaming, Social media platforms and Stock Trading system.

In the recent years the need of the recommendation system has increased a lot because recommendation system has been using in almost every application or platform and it is also playing a vital role by providing a products similar to users interest. As there is a saying nothing comes free, so with the extensive usages of recommendation systems and the advancement in recommendation system there is a one major problem which is still need to be solved and that is Cold start problem.

Get quality help now

Doctor Jennifer

Verified writer

Proficient in: Technology

5 (893)

“ Thank you so much for accepting my assignment the night before it was due. I look forward to working with you moving forward ”

+84 relevant experts are online

Hire writer

As we mentioned earlier that recommendation system has been in almost every where to meet the interests of the customers but our main focus is dealing with cold start problem in online streaming recommendation systems like Netflix, Hulu and Amazon prime videos etc. Because many people now prefers to watch entertainment sitting in their home and they love to watch the content if it had been offered to them according to their interest.

Research Question

Cold start problem in recommendation system are those situations in which recommendation system or model does not have a ratings or clicks related to a specific user or related to a specific items.

So to make a prediction similar to a users interest models is not that simple and it is needed an additional information like users’ geographical location, users’ age and their genders etc or the first interviews during the start of the recommendation system. Our main focus and our research question will be the dealing with such a cold start problem in online streaming recommendation system and try to minimize the cold start problem in online streaming recommendation systems. Providing recommendations to users specially in the beginning which are not similar to user’s choice or interest may ends up in users’ frustration and which will eventually lead the user to stop using recommender system.

Objectives

One of our main objective is to study the already available literature work on dealing with the cold start problem in online recommendation systems, because it will help us to focus on the right track. Many people in industry already done a research on dealing with cold start problem and they have shared a quite good methods to deal or to minimize the cold start problem in online streaming recommendations. So after reading the literature work, our another objective is to propose a technique by enhancing the existing techniques in terms of accuracy which makes the model works better than the existing online streaming models.

After implementing a model, our objective is to evaluate the model to check either the model or the proposed technique is really enhanced, gives a better results or not. we will test a model on the available training and testing data sets. After testing we will do improvements or changes in a techniques according to requirement or the feedbacks. This evaluation could be done in a lab or could be done in a real environment. And once we are satisfied with the model, we can then deploy into real world scenario.

Methodology

In order to deal with a cold start problem in recommendation systems in online streaming applications, literature review and research would be conducted. The goal is to propose a new method or it would be better if we put in this way, that enhancing the already mentioned techniques[1][2] just to improve the accuracy of recommendation systems in online streaming. One of the way to deal with the new user cold start problem is to use similarity measure. In the paper[1] authors has given a method or proposed a new similarity measure which can be used in collaborating filtering to deal with Cold start problem related to only new user. Basically authors presented a heuristic approach with is known as PIP (ProximityImpactPopularity) measure. As clear from its name this similarity measure based or composed on three factors which is Proximity, Impact, and Popularity. The similarity they have introduced is compose on a number of goals.

The first goal which authors has introduced in the paper [1] is that measure need to use domain knowledge of the data because authors believed it will make it more effective. Another goal which authors have defined is that this measure can also be used in comparable results to already existing measures in a non-cold start scenarios. The third and last goal which authors have defined is that this measure can be used in already existing recommendation system, during implementation it just replacing the measure. Another technique or factor which will be helpful to deal or to minimize the cold start problem is Geographical factor.

Geographical factor impacts a lot on initial recommendation or in a scenario like cold start. Using this factor helps in this way, many people which belongs to a same geographical usually have similar kind of interest in different genre. Because in different regions of the world different trends are going on. In the paper[2] authors have presented the similar approach to deal with the cold start problem. In the paper authors suggested an approach which is based on demographic data. On the basis of this demographic data they will provide recommendations in a cold start scenario. The recommendations which will be provided to user will be based on different demographic attributes rather than the users’ rating.

Basically the demographic-based approach which authors has proposed in the paper[2] based on three stages. First stage is the data input, second stage is the similarity calculation and the third or last stage is recommendation calculation. First stage which is the data input, takes the demographic data of the new users which are basically the target users. These target users are the users who needs the recommendations.

Besides the target users first stage also takes into account of rating and demographic data of the others users. Second stage which is known as similarity calculation stage basically use users’ demographic data to acquire a more users which have demographic data in common or similar with respect to the target users so it can form a neighborhood. Finally the third and the last stage which is recommendation calculation stage acquire items which is rated by neighborhood users but in a positive-rated term, so can on the basis of this it can suggest a recommendations to target or new users.

After mentioning the two techniques in the previous paragraphs, we are proposing an approach in which we will use a hybrid (methods which are mentioned on [1][2] ) and classification technique which will be based on the geographical factor. And this geographical factor will use all the users’ geographical attributes like users’ age, users’ gender, users’ age and occupation. The idea of this technique is to use the similarity measure which authors have introduced in paper[1] which is as PIP (ProximityImpactPopularity) measure with an other approach which is mentioned in the paper[2]. The idea of this technique works in a very simple way. This technique basically composed on two steps. As we know in paper[1] there is only one similarity measure but if we look at this paper [2] their proposed technique based on 3 stages which are basically data input, similarity measure and recommendation calculation. As the authors of the PIP mentioned in their paper that this measure can be used in other recommendations system easily. So basically we will use this PIP measure as a second stage.

And the first stage (data input ) and the last stage or the third stage will remain the same. Once the model has learned the this hybrid technique, the next step of this technique is to combine this model with the machine learning classification’ technique. This classification method will basically keep the data of the users and the recommended items in a separate classes. This classification technique or step will use different classification algorithms to fit the situation or to handle the cold start problem. Because after executing or training the model few times, it will automatically classify the user class among the region and will provide the similar recommendation to user from a number of items classes.

The main goal behind this introducing the approach is to minimize the cold start problem. Once we are done with the implementation step then we need to move towards the evaluation step. Evaluation could be done in different ways but we need to look which evaluation ways suits our model better. There are two ways to evaluate the results of the model we proposed. First way is to test the accuracy or performance of the model in a controlled environment / lab and notice how this recommendation system is dealing with the cold start problem and on the basis of the results we will try to improve the model’s performance or accuracy as required to minimize the cold start problem. This controlled environment consists on a number of hired people who will test the model in a cold start scenario.

And the second method to evaluate the model is to deploy the model in small area or region initially to see the how this model works and then once we got some values then to test the accuracy of the model and see how this model deals with the problem in real environment. This second method could be a little risky because we have to deploy the system in a real world and nobody knows how the users of the system will behave. And on the basis of the results, we will improved the model. If the model starts giving a better accuracy or performance then eventually it will be deployed in the real environment and with the passage of time will try to improve the model on the basis of feed backs.

State of the Art

To my knowledge different work has been carried out in order to deal with the cold start problem in recommendation systems. As we know in the paper [3] authors has presented a collaborating filtering approach just to deal with this problem. The similarity measure approach which they have introduced, is based on neural learning. Besides similarity measure deep learning techniques are also using in order to deal with the cold start problem. As in the paper [4], authors has presented a deep learning method to handle the cold start problem. Another method or technique which authors have introduced in the paper[5] is matrix factorization to deal with the cold start problem. Basically their approach will solve the problem of how first interview has been constructed in order to get familiar with the users and the items. Some of this work is direct connected or related to our work and some of it gives us a understanding that this problem will behave in a different situations.

Project Plan

In our project plan, we have decided to go step by step. To achieve a final or bigger goal, we have divided that one big goal into a number of steps or small goals to achieve that bigger or final goal. We have set the every goal (small) with a specific deadline to achieve it. In this way we will keep the track of our performance and will remain on schedule too to deliver the final goal or to achieve the bigger goal in the end. It is quite common in software processes to solve a some particular problem, first we need to understand that what is the problem and what this problem actually states. So that is why first we will have a proper research on the problem and once we comprehend the problem completely then after that we can take a stance or make a statement that this problem need to solved or need to be addressed and why it is so important.

After that we have spare some time for literature research.In literature research we will come to know the contributions of others people in solving the same problem. Once we are done with the literature research, then we will have to conclude the finding of the literature review because as mentioned in previous section that existing techniques will be used in our project. After these steps we are good to go for implementation because we will have all information we needed to start a implementation. Once we are done with the implementation we need to evaluate our model either in a lab or in a real environment to test the accuracy of the recommendation system. Later we have spare a time to make an improvements or changes according to the results or feed backs. And the last three milestones has been given to documentation, which includes the first draft of the paper, revision version and the final one.

Relevance, Originality, Benefits

As mentioned in methodology section, the introduced technique will use these mentioned techniques [1] [2] as part of the approach. The difference of the introduced approach is that not only it enhancing the previous methods but it also introducing the machine learning classification technique as a next step. This classification technique will help the recommendation system to deal cold start problem by classify the different regions, different users, different areas and the recommended items with respect to these geographical factor. The solution of this approach that it will deal with the Cold start problem in online streaming and will provides a better results and accuracy to users.

Preconditions

The tests to evaluate the model will be conducted on MovieLens and Netflix data sets respectively. The data set is already publicly available on a Kaggle site, so with respect to that we do not any need permission to get access to data sets. While conducting this research, we do not need any special software licenses. To complete this project from start to end, we do not need any funding and neither we have mentioned or asked for this requirement. During the literature review or research, intellectual property will respected and will be given proper credit where it is due.