Adamas University Students: Building a Recommendation System Mini Project During an Internship

Yes, you can build a credible recommendation system mini project in a 14-day internship, provided you stage it: a popularity baseline in the first days, an item-based method next, then user-based or matrix-factorisation collaborative filtering, with an evaluation checkpoint after each stage. What makes the project convincing is not the fanciest algorithm but a clear comparison showing each stage beats the previous one on metrics you chose in advance.

This guide gives you the full staged plan, a checkpoint table, a rubric for judging recommendation quality, an illustrative worked example, and a short section for Adamas University students on the internship paperwork that surrounds a project like this.

What a Recommendation Mini Project Should Deliver in Two Weeks

Recommenders are a popular AI project because the data is easy to find, the idea is intuitive, and the results can be measured. A two-week scope forces discipline. Aim for these deliverables:

• A cleaned interaction dataset (user, item, rating or implicit signal, optional timestamp).
• A popularity baseline that every later model must beat.
• At least one collaborative filtering model (item-based neighbourhood is the safest choice; matrix factorisation is a good stretch).
• An evaluation script with a fixed train/test split and at least three metrics.
• A short README or report with a results table, a few qualitative examples, and honest limitations.

Do not try to build a web app, a deep neural recommender and a hybrid system together. One well-evaluated pipeline is worth more than three half-working demos.

Choosing a Dataset You Can Finish With

Pick a public dataset small enough to run on a laptop or free notebook environment. Movie ratings datasets of around 100,000 ratings are the standard learning choice. Book ratings, anime ratings or music listening logs also work. Check the licence and terms on the dataset’s own page before you publish anything.

Use this quick filter before committing:

1. Does every row have a user ID, an item ID and some signal of preference?
2. Does each user have enough history? Aim to keep users with at least 5 interactions so you can hold some out for testing.
3. Is the matrix manageable? A dense in-memory table of users by items becomes painful past a few tens of thousands of each, so prefer sparse matrices.
4. Does it include item titles or categories so you can sanity-check recommendations by eye?

The 14-Day Staged Build Plan, Baseline to Collaborative Filtering

Stage 1: Popularity baseline (Days 4-5)

Popularity is not a throwaway step. It is the benchmark that tells you whether personalisation is adding anything. Build two versions: most-interacted items, and highest average rating with a minimum count (for example, at least 20 ratings) so one five-star vote does not top the chart. Remove items the user has already seen from each user’s list, otherwise your metrics will be artificially distorted.

Stage 2: Item-based collaborative filtering (Days 6-8)

Build a sparse user-item matrix, compute item-to-item cosine similarity, and score a candidate item for a user by summing similarities to the items the user liked. Keep only the top-k neighbours per item (try 20, 50 and 100) to save memory and reduce noise. Item-based methods are easy to explain in a viva: “people who liked these items also liked this one.”

Stage 3: User-based or latent factors (Days 9-10)

Pick one. User-user filtering is conceptually simple but slower as users grow. Truncated SVD or ALS captures hidden taste dimensions and often improves ranking, but you must explain what the factors mean and how you chose the number of factors. Do not add both unless the schedule is clearly ahead.

Evaluation Checkpoints That Keep the Project Honest

Which metrics to report

• Precision@K and Recall@K: of the K items shown, how many did the user actually interact with in the held-out data, and how many of the held-out items were found.
• NDCG@K or MAP@K: rewards placing relevant items higher in the list.
• Coverage: share of the catalogue that ever gets recommended. Popularity baselines score badly here.
• RMSE or MAE: only if you predict explicit ratings. Rating error alone says little about the quality of a top-N list.

Use a time-based or leave-last-out split when timestamps exist; random splits can let the model peek at the future. Fix your seed, evaluate on the same users for every model, and tune only on validation data.

Rubric for Judging Recommendation Quality

Numbers alone miss problems you can see by reading the lists. Score your final model on this rubric, 0 to 3 per row (0 = absent, 1 = weak, 2 = adequate, 3 = strong), for a maximum of 24.

As a rough reading, 19-24 suggests a project ready to present, 13-18 means one or two criteria need rework, and below 13 means revisit the evaluation before adding more models. These bands are a self-assessment aid, not an official grading scheme.

Worked Example: Reading One Illustrative Results Table

The numbers below are invented for illustration only. They are not output from any real run. Suppose a student uses a 100,000-rating movie dataset with a leave-last-out split and K = 10.

How to read it: both personalised models beat popularity on recall and ranking, so Checkpoint 2 passes. Coverage rises sharply, showing the lists are less repetitive. The gap between item-based CF and matrix factorisation is small, so a careful write-up would say the improvement may be within noise unless confirmed across several seeds or user samples. The student should then pick three users, print their watch history next to their top-10 lists, and note any odd results, such as a children’s film recommended to a user who only rated thrillers. That qualitative note earns marks under the explainability and honesty rows of the rubric.

Pitfalls That Quietly Break Recommender Projects

• Data leakage: computing similarities on the full dataset before splitting inflates test scores.
• Recommending seen items: always mask items already in the user’s training history.
• Tuning on the test set: use validation data for choices like k and number of factors, then touch the test set once.
• Ignoring popularity bias: a model that mostly echoes the baseline can still look decent on accuracy.
• Dense matrices: converting a large sparse matrix to dense will exhaust memory; keep it sparse.
• Claiming real-world impact: offline metrics do not prove users would click or enjoy the results.

Documenting the Project So It Counts as Evidence of Work

Keep a dated daily log from Day 1: what you tried, what failed, what the metric moved to. Commit code at least daily so your repository history shows steady progress. Your final report should contain the problem statement, dataset description, split method, model descriptions, the results table, rubric self-score, and limitations. If you are looking for a structured way to practise this, the free two-week online internship from CareerSwami offers 26 tracks, a free offer letter and a free completion certificate after successful completion, and a project like this fits naturally into a 14-day window.

Adamas University Students: Aligning an Online Recommender Project With CDC Internship Rules

Adamas students should treat the technical plan above as separate from the university’s paperwork, because the university has its own internship procedure. The 2024-25 Placement & Internship Policy applies to all registered students and is administered centrally by the Career Development Centre (CDC). The same document says it explicitly recognises virtual internships, so an online project is not excluded in principle.

Several points in that policy matter if you want any university recognition. It states an internship period of a minimum of 3 weeks and a maximum of 8 weeks during summer vacation, which is longer than a 14-day programme. It also says that for credit, students need 100% attendance, a daily log book countersigned by a supervisor, a formal report and a presentation to a committee formed by the school Deans. A CDC reference letter and NOC are required before starting any internship, and without them a student is marked absent. Note that the document still labels its rules as 2021-22, so check whether a newer version applies.

For self-arranged internships, the CDC handbook says the NOC request goes through your HOD to internship@adamasuniversity.ac.in and needs the company’s offer letter. It adds that NOCs are issued only for internships within the approved academic calendar. Programme structures also differ: the B.Tech CSE (Cloud Computing) structure lists a 2-credit Summer Internship after semester 6 and a 3-credit Minor Project in semester 7. Whether a 14-day online programme could satisfy any of these requirements is not stated in the sources, so do not assume it does. A recommender mini project may be useful as portfolio work regardless.

Verification steps for Adamas students before starting a recommender internship

1. Email or visit your HOD and ask in writing whether a 14-day online internship can count toward any credit, given the policy’s 3-week minimum.
2. Request the CDC reference letter and NOC before the first day, and keep the reply.
3. Ask CDC at internship@adamasuniversity.ac.in whether the 2024-25 policy clauses are current, since the document still carries 2021-22 labels.
4. Check your own programme’s course structure for the internship or minor project credit, semester and evaluation method.
5. Confirm the current academic calendar dates on the university’s notice board before fixing your internship window.
6. Ask whether your log book, report and presentation format is prescribed, and obtain the template if so.

Frequently Asked Questions

Is a popularity baseline really necessary in a recommender project?

Yes. Without it you cannot show that personalisation helps. A simple most-popular list is surprisingly hard to beat on sparse data, so reporting it protects you from overclaiming.

Should I use item-based or user-based collaborative filtering first?

Start with item-based. Item similarities are usually more stable, easier to precompute, and simpler to explain. Add user-based or matrix factorisation only after the item-based result is evaluated.

How do I split data when I have timestamps?

Hold out each user’s most recent interactions as validation and test, and train on earlier ones. This mimics real use, where you predict the future from the past, and avoids leakage that random splits can cause.

Which metric matters most for a top-N recommender?

Ranking metrics such as NDCG@K, Recall@K and MAP@K are better matches than RMSE, because users see a short list rather than a predicted rating. Report coverage alongside them.

How should I handle new users with no history?

Fall back to the popularity list, or popular items within a category the user selects. Report how many test users triggered the fallback so readers understand how often it matters.

Can I include this project in my resume if the internship was only 14 days?

Yes, describe it accurately: dataset, methods, metrics and your own contribution, with the repository link. Avoid implying production use or measured business impact, since the evaluation is offline.

Will a CareerSwami recommender project count towards university requirements at Adamas?

That is not stated in the sources reviewed, and CareerSwami makes no such claim. Ask your HOD and the CDC in writing whether a short online internship can count, and use the project as portfolio evidence either way.

Related Artificial Intelligence guides

• CMR University Students: Presenting an AI Internship Project to Non-Technical Managers
• IIMT University Students: Choosing a Dataset for an AI Internship and Checking Its Licence
• Manipal University Jaipur Students: Cracking the AI Internship Behavioural Interview with STAR Stories
• All Artificial Intelligence internship guides